跳到正文
只看一手
10月20日周二
  1. GitHub Blog · AI & ML

    How to continue a GitHub Copilot app session in VS Code

    The agent has finished the bulk of your change. Now you want to inspect one implementation detail, adjust a few lines by hand, and run the project locally before you ship it. But the work is sitting in the GitHub Copilot app … and your editor is where you review diffs, debug code, and make final adjustments. Do you have to choose between staying in the session and starting over in Visual Studio Code? You don’t. You can open the Copilot app session directly in VS Code and continue with its conversation, code changes, and working context intact. That means Copilot can handle the repetitive or time-consuming work, then hand control back to you for closer review, hands-on editing, and local testing. In this tutorial, you’ll learn how to: Open a Copilot app session in VS Code. Review the agent’s changes line by line. Continue the same conversation from your editor. Modify and test the code locally. Push your updates to the pull request associated with the session. Prerequisites Before you begin, make sure you have: The GitHub Copilot app installed and signed in. Visual Studio Code installed and connected to the same GitHub account. A Copilot app session with code changes you want to review or continue. Permission to work in and push changes to the session’s repository. This tutorial assumes you are already comfortable navigating VS Code, running commands in its integrated terminal, and working with Git branches and pull requests. 1. Start with a session in the Copilot app Open the session you want to continue. Before handing it off, take a moment to review the task and the work Copilot completed. A session contains more than a collection of modified files. It also records the task, the plan, and the conversation that produced those changes. That context is important when you want to understand why Copilot chose a particular implementation or ask it to revise part of the work. You do not need to summarize that history or copy it into a new chat before moving to VS Code. 2. Open the session in VS Code From the session, select Open in VS Code . If the app presents more than one supported editor or VS Code installation, choose the one you want to use. VS Code opens the project using the worktree associated with the session. It also displays the existing session in the Copilot sidebar. That gives you three connected pieces of the original work: The session context: The task, plan, and conversation remain available. The diff: The files and lines changed by the agent are ready for review. The worktree: VS Code opens the same isolated working tree the session used. You are continuing the existing session, not reconstructing it in a new editor window. 3. Review the agent’s changes Before editing, inspect the diff in VS Code. Open the Source Control view to see the changed files, then select a file to review its changes line by line. Review the implementation as you would any other proposed change. In particular, check that: The code solves the requested problem. The implementation follows the repository’s conventions. Error states and edge cases are handled appropriately. Tests cover the behavior that changed. No unrelated files or behaviors were modified. The preserved conversation can help explain how Copilot arrived at the implementation. If something is unclear, continue prompting Copilot in the sidebar. For example: Explain why you changed this function instead of adding a new helper. Identify any behavior this could affect. You can also ask Copilot to make another change. Because the session and worktree remain connected, you do not have to describe the task from the beginning. 4. Make a hands-on edit Once you understand the proposed changes, edit any file directly in VS Code. This is useful for adjustments that are faster or easier to make by hand, such as: Refining a label or error message. Renaming a variable for clarity. Adjusting layout or styling. Adding a missing test case. Applying a repository-specific convention. You are editing the session’s existing worktree, so your manual changes build directly on the agent’s work. You do not need to copy files between directories or manually re-create the branch. 5. Run and verify the project locally Use the VS Code integrated terminal to run the project’s normal validation workflow. The exact commands depend on the repository, but some web app examples may include: npm install npm test npm run lint npm run dev Treat agent-generated changes the same way you would treat code written by a person: run the relevant tests, inspect failures, and verify the behavior yourself. If the development server prints a localhost URL, open it to test the change in a browser. When supported by your VS Code setup, you can open the URL in VS Code’s integrated browser and inspect the result without leaving the editor. For a user-interface change, verify both the requested behavior and nearby states. For example, check the modified component at different viewport sizes and confirm that loading, empty, and error states still work. 6. Continue with Copilot or keep editing yourself At any point, you can choose the interaction that best fits the next step: Prompt Copilot when the change requires repository-wide analysis or repetitive edits. Edit directly when you know the exact change you want. Use both approaches iteratively as you test and refine the implementation. For example, after making a manual UI adjustment, you might ask Copilot: Add a test for the updated label and run the relevant test file. Do not modify unrelated components. Because you are still in the original session, Copilot retains the context needed to interpret “the updated label.” 7. Commit and push the finished changes When the implementation is ready, review the final diff one more time. Then commit and push the changes using the Source Control view or the integrated terminal: git status git add git commit -m "Refine session changes" git push Use your repository’s normal branch, commit, review, and approval practices. Because VS Code opened the worktree and branch associated with the session, pushing the branch updates the pull request already tracking that work. There is no need to copy the finished files into another checkout or reconnect the work to the pull request manually. A continuous workflow from agent to editor Opening a Copilot app session in VS Code gives you a clean division of labor without splitting the work itself. Copilot can take on a larger or repetitive implementation, while you can move into the editor for detailed review, targeted changes, local testing, and final polish. The session, diff, and worktree move with you. Start by directing the work in the Copilot app, switch to VS Code whenever you want hands-on control, and push the finished result through the same pull request workflow. Ready to try it? Download the GitHub Copilot app , start a session, and open it in VS Code when you are ready to continue coding. The post How to continue a GitHub Copilot app session in VS Code appeared first on The GitHub Blog .

  2. arXiv · 渲染、神经场与导出

    Counting Contributions, Not Views: Occlusion-Aware Densification for 3D Gaussian Splatting

    3D Gaussian Splatting grows a dense scene from a sparse initialization through adaptive density control (ADC), driven by per-Gaussian statistics gathered during rendering. These statistics pass a frustum-based visibility filter blind to occlusion: fully occluded Gaussians still count as observed, diluting the averaged gradient signal and suppressing densification precisely where geometry is occluded. We correct this with an occlusion-aware filter that restricts densification statistics to views in which a Gaussian actually contributes, and replace the criterion's raw view-count normalizer with an average-contribution one. Integrated into a StopThePop-based framework, the two together improve every image-quality metric over the occlusion-blind baseline on occlusion-heavy scenes (up to 6% lower LPIPS), at less than half the Gaussians a view-count normalizer requires. Under the longer schedules these large scenes require, the count stays bounded where most criteria diverge, and on smaller benchmarks our method stays competitive at a fraction of the competitors' count.

  3. arXiv · 创作 Agent 与装配

    LVMem: Image-Augmented Memory for Long-Horizon Question Answering

    Reliable language-guided perception is essential for embodied agents, where questions about objects, events, and their locations must be answered from long observation histories. Unlike conventional visual question answering, realistic robotic settings require reasoning over extended temporal horizons and persistent scene memory. We build upon the view-based memory paradigm and propose an efficient framework for long-horizon spatio-temporal question answering. While existing view-based systems typically store either text captions or visual embeddings, we extend the ReMEmbR-style memory with image paths that link each stored description to its original visual content. To make the stored annotations more informative, we replace uniform sampling during memory construction with an optical-flow-based frame selection strategy (PSFR) and store an object-centric descriptive scene representation instead of plain captions. We further equip a VLM-based agent with an adaptive retrieval tool that returns the image associated with the most relevant caption, allowing the agent to decide when visual inspection is needed. On the OC-NaVQA benchmark, LVMem outperforms ReMEmbR on binary and positional metrics with both backbones and matches or exceeds a 4D scene graph DAAAM baseline on binary accuracy, while operating on a lightweight view-based memory that requires no explicit graph construction.

  4. arXiv · 资产数据集与许可

    VideoMaMa++: Temporally Consistent Video Matting via Preserve-and-Refine Token

    Diffusion-based video matting has recently emerged as a leading approach, leveraging video diffusion priors from internet-scale pretraining to achieve strong zero-shot generalization from limited synthetic data. However, the heavy compute of these diffusion backbones limits inference to a short chunk of just 6 to 24 frames, so even ordinary videos span multiple chunks. Because each chunk is matted independently from coarse mask guidance, existing methods fail to maintain consistency across chunk boundaries, producing flickering edges and discontinuous transparency. We address this with a mask-guided diffusion-matting framework built on two ideas. The RGB and mask conditioning are represented as separate token streams that interact through full spatio-temporal self-attention, letting the model adaptively combine fine RGB texture with coarse mask layout rather than committing to a fixed channel-wise fusion. On top of this, we propose a novel pair of learnable embeddings, the preserve token and the refine token, that act as a per-frame conditioning interface and enable a sliding-window inference scheme in which already-generated mattes are propagated across chunk boundaries for local temporal consistency, complemented by a globally shared reference anchor at every chunk to prevent error accumulation in long sequences without any additional training. Across standard video matting benchmarks and in-the-wild long footage, our framework outperforms prior diffusion-based methods in both fine-detail accuracy and temporal consistency, and uniquely maintains consistency across chunk boundaries on long videos.

  5. arXiv · 三维交互与编辑

    VitronEdit: Universal Visual Editing across Images, Videos, and 3D Scenes

    Recent advances in unified visual editing have begun to consolidate image and video editing within a single model, while 3D scene editing remains largely separate and constrained by scarce paired training data. Given the rich image and video editing supervision now available and the common editing semantics shared across modalities, a natural question is whether this knowledge can be transferred to data-scarce 3D editing. We present VitronEdit, a universal framework for image, video, and 3D scene editing. VitronEdit adopts a shared sequence representation and progressive multimodal training to transfer editing knowledge across modalities. We further introduce geometry-aware post-training, which leverages a pretrained geometry reconstruction model to provide depth and camera-pose supervision. Surprisingly, applying geometric supervision only to 3D samples yields the best overall performance, while image and video editing also improve despite receiving no direct geometric supervision, suggesting that geometric regularization transfers through the shared backbone. Our analysis further shows that image and video training mainly transfers editing semantics to 3D, whereas 3D supervision primarily improves multi-view consistency. The project page: https://vitron-llm.github.io/VitronEdit/.

  6. arXiv · 创作 Agent 与装配

    Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience

    World models enable embodied agents to anticipate action consequences, yet jointly modeling locomotion and manipulation across diverse embodiments remains challenging due to heterogeneous action spaces and limited multimodal supervision. We present Ropedia W0, the first egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. We curate Xp-Data, comprising 10k hours of egocentric human data and 3k hours of robot data with unified multimodal annotations and systematic quality control. A unified kinematic representation encodes locomotion through camera poses and manipulation through projected skeletons, with robot skeletons combining embodiment-specific forward kinematics with extensive manual keypoint annotation and correction. We also introduce Xp-Teleop, a long-horizon loco-manipulation benchmark based on high-quality dexterous humanoid teleoperation data. Evaluations demonstrate state-of-the-art performance across five cross-embodiment benchmarks (DreamDojo, $Ψ_0$, $ω$-HOME, Xp-Teleop, and AgiBot) and on held-out human egocentric video, from zero-shot inference to few-shot and full-data post-training. Scaling human pre-training data improves performance on both human and robot evaluations. After distillation, our model supports real-time, minute-long rollouts, enabling live VR interaction. Applications in policy evaluation, model-based planning, and world-action modeling further demonstrate the model's utility for embodied decision-making.

  7. arXiv · 创作 Agent 与装配

    PVBench: Quantifying Physical Plausibility in Video Generation with Object-Level Physical Measurements

    Recent video generation models produce visually compelling outputs, yet their physical plausibility remains poorly measured. Physical plausibility asks whether generated objects obey geometry, kinematics, dynamics, and conservation laws. Existing benchmarks rely on coarse visual-language model judgments or target embodied world models rather than video generation itself. We introduce PVBench, a physical video benchmark that quantitatively measures the plausibility of generated videos across 46 dimensions organized into seven domains, spanning: i) instruction and semantic fidelity, ii) behavioral and causal consistency, iii) perceptual and temporal quality, iv) geometric and spatial consistency, v) material properties and field states, vi) kinematic and spatiotemporal persistence, vii) and mechanical plausibility with conservation laws. PVBench comprises 600 carefully crafted prompts, of which 30% involve physical embodied agents such as robot arms, humanoids, quadrupeds, and multi-agent systems. For challenging physically tractable dimensions, it pairs each with specialized vision models that derive verifiable, physically grounded measurements, such as object-level 6D pose, metric scale, surface normals, kinematics, and numerical conservation-law residuals. It thus avoids relying solely on subjective scoring, while supporting both text-to-video and image-to-video generation. Finally, we validate the agreement between automatic metrics and human judgments through human evaluation. Meanwhile, we evaluate mainstream open-source and closed-source models and present several insights. By releasing the benchmark, prompts, and calibration protocol, we aim to foster physically grounded video generation.

10月19日周一
  1. arXiv · 创作 Agent 与装配

    BetaMem: Learning to Remember a Changing World for Embodied Agents

    Long-horizon embodied agents face temporal confusion and incomplete or distorted records, recurring errors from fixed memory rules, and growing storage and retrieval costs. We present BetaMem, a framework for building and evolving traceable long-term memory through six linked layers: working, spatial, entity, episodic, experiential, and procedural memory. A dual-level write contract constrains operation executability and source support for content: supervised fine-tuning learns compliant memory operations, while reinforcement learning combines repeated sampling with teacher verification of the original observations to reward complete, accurate records and penalize omissions, distortion, and unsupported content. We formulate failure-driven policy recursive self-improvement through diagnosis along the causal sequence of memory operations and local rule repairs gated by independent replay. Atomic tools frame storage, updates, retrieval, and forgetting as cost-utility allocation under storage and read budgets. Using the open-source Qwen3.5-9B model for memory generation, BetaMem achieves overall scores of 48.08% on M3-Bench-robot, 92.08% on LoCoMo, 72.6% on OpenEQA EM-EQA, and 62.4% on EgoLifeQA, the highest among the listed memory methods under their respective evaluation protocols. Training the model for memory generation raises the contract pass rate from 50.9% to 86.2% while reducing output tokens per stored action by 47.3%. A separate Qwen3-VL-8B-Instruct evaluation on EMem-Bench achieves an average success rate of 38.4% and an average error recurrence rate of 24.9%. We perform best among evaluated methods using comparably sized open-source models. Category results show that we have advantages in integrating information across observations, recalling events over long histories, and inferring personal habits.

  2. GitHub · WhiteTowerAI/ohmygame

    WhiteTowerAI/ohmygame 发布 OhMyGame v0.0.0-beta.4.10

    ### Added - Use installed Claude Code, Codex, and OpenCode CLIs in conversations, with their own accounts and models, native reasoning controls, and OhMyGame's tools, checkpoints, and Undo. ([#299](https://github.com/WhiteTowerAI/ohmygame/pull/299) by @dogxii) - Manage coding agents in Settings → Agents, choose Standard or Full access permissions, and view reported Claude Code and Codex account usage limits. ([#302](https://github.com/WhiteTowerAI/ohmygame/pull/302) by @dogxii, [#304](https://github.com/WhiteTowerAI/ohmygame/pull/304) by @dogxii, [#305](https://github.com/WhiteTowerAI/ohmygame/pull/305) by @dogxii) - A Visual Novel template with editable scripts, choices, save slots, Auto, Skip, Log, and Settings, plus cross-Scene music and richer Main menu controls. ([#284](https://github.com/WhiteTowerAI/ohmygame/pull/284) by @dihuangdh) - Trim videos in the Library and on Asset Canvas, preview the selected range, and save it as a new video while keeping the original. ([#265](https://github.com/WhiteTowerAI/ohmygame/pull/265) by @dihuangdh) - Trim a Scene's background video without changing the original asset or other Scenes, with precise time controls and Undo/Redo. ([#312](https://github.com/WhiteTowerAI/ohmygame/pull/312) by @82228) - Generate sound effects and instrumental music with Volcengine Seed Audio. A compatible connected provider is required. ([#288](https://github.com/WhiteTowerAI/ohmygame/pull/288) by @82228) - Open a project's workspace folder from recent-project and Projects menus in the desktop app, replacing the obsolete Game design shortcut. ([#315](https://github.com/WhiteTowerAI/ohmygame/pull/315) by @dogxii) ### Changed - New Interactive Stories run as independent web pages, with a project-owned runtime, persistent saves, and static publishing and HTML export. Existing stories keep working and can upgrade with an option to undo the upgrade. ([#284](https://github.com/WhiteTowerAI/ohmygame/pull/284) by @dihuangdh) - New projects can be named automatically from the first meaningful request; stable icon and color placeholders make projects easier to recognize. ([#291](https://github.com/WhiteTowerAI/ohmygame/pull/291) by @dogxii) - Checkpoints respect workspace `.gitignore` rules and prune unreferenced snapshots without removing recorded turn history. ([#280](https://github.com/WhiteTowerAI/ohmygame/pull/280) by @xuhongbo) - The Windows setup wizard lets users choose the installation directory and current-user or all-users installation, preserving existing upgrade settings. ([#306](https://github.com/WhiteTowerAI/ohmygame/pull/306) by @82228) - Seedance video generation includes audio. ([#271](https://github.com/WhiteTowerAI/ohmygame/pull/271) by @82228) - Release notes credit the authors of linked contributions. ([#283](https://github.com/WhiteTowerAI/ohmygame/pull/283) by @xuhongbo) ### Fixed - Codex's macOS command sandbox can read the system OpenSSL configuration required by Node/npm when installing project dependencies. ([#321](https://github.com/WhiteTowerAI/ohmygame/pull/321) by @dihuangdh) - Compatible coding-agent CLIs are checked for required capabilities instead of rejected solely by version; local compatibility and startup failures have clearer diagnostics. ([#317](https://github.com/WhiteTowerAI/ohmygame/pull/317) by @dogxii) - Codex-generated images appear in the conversation and are available to the project. ([#318](https://github.com/WhiteTowerAI/ohmygame/pull/318) by @dihuangdh) - Preview servers keep running during checkpoint snapshots, and file reads remain available while writes wait for a consistent snapshot. ([#268](https://github.com/WhiteTowerAI/ohmygame/pull/268) by @dihuangdh, [#273](https://github.com/WhiteTowerAI/ohmygame/pull/273) by @dihuangdh) - Static video previews support seeking, and HEAD requests ignore byte ranges. ([#272](https://github.com/WhiteTowerAI/ohmygame/pull/272) by @dihuangdh, [#300](https://github.com/WhiteTowerAI/ohmygame/pull/300) by @jsdhwfmax) - AI plugin authoring validates required Skill metadata and supports configuring and testing drafts before installation. ([#277](https://github.com/WhiteTowerAI/ohmygame/pull/277) by @dogxii) - Git Changes compares files removed from the index against HEAD, and project exports avoid asset-name collisions on case-insensitive file systems. ([#281](https://github.com/WhiteTowerAI/ohmygame/pull/281) by @jsdhwfmax, [#269](https://github.com/WhiteTowerAI/ohmygame/pull/269) by @jsdhwfmax) - Pages keep their horizontal alignment when scrollbars appear, and the plan mode label no longer wraps in narrow composer toolbars. ([#303](https://github.com/WhiteTowerAI/ohmygame/pull/303) by @dogxii, [#308](https://github.com/WhiteTowerAI/ohmygame/pull/308) by @dogxii) - The extra-high reasoning level is translated in the Simplified Chinese UI. ([#282](https://github.com/WhiteTowerAI/ohmygame/pull/282) by @dogxii) **Full Changelog**: https://github.com/WhiteTowerAI/ohmygame/compare/v0.0.0-beta.4.9...v0.0.0-beta.4.10

  3. arXiv · 资产数据集与许可

    Local data augmentation for imbalanced binary classification

    Facing imbalanced binary classification, we aim at maximizing the recall while limiting the proportion of false detections in the minority class. For doing that, the present work studies the effect of several rebalancing strategies, from the uniform subsampling of the observations from the minority class, to a more involved local data augmentation scheme. To be more specific, we establish that only rebalancing class proportions cannot modify the classification performance, whereas local data augmentation is proved to have an impact on the final results. The local augmentation scheme is introduced by means of a two-step procedure. The first step randomly subsamples anchor points among the training observations. These anchor points are used in the second step, for drawing new augmentation data following an ''augmentation kernel'' to be specified. This enables us to decouple the impact of the anchor points subsampling from the one of drawing augmentation data accounting (or not) for the local geometry of the support. The experiments on synthetic data illustrate the practical performance of several strategies for subsampling anchor points combined with either a Gaussian or a diffusion-based augmentation kernel. Compared to SMOTE-like data augmentation procedures, a better performance is achieved when subsampling in the close neighborhood to the classification boundary combined with a Gaussian augmentation kernel. Furthermore in specific settings where the geometric structure of the support matters, the experimental results show that using a diffusion-based augmentation kernel outperforms the Gaussian one.

  4. arXiv · 角色动作与重定向

    Streaming Multi-Track Timeline Control for 3D Human Motion Generation

    Text-driven human motion generation has advanced substantially, yet most methods assume instructions are available before synthesis. Interactive applications require responding to new instructions while continuing ongoing actions, such as answering a phone while walking. Existing approaches address streaming generation or simultaneous composition without explicitly combining streaming instruction arrival with independently timed, overlapping actions. We introduce streaming multi-track timeline control and propose TimelineControl to incorporate new instructions alongside ongoing actions. Interval-aware conditioning preserves instruction timing, while causal part-structured representations and part-aware denoising coordinate concurrent actions across body regions. We also construct TimelineMotion, a dataset with overlapping instruction intervals and body-part annotations. Experiments on TimelineMotion and MTT demonstrate improved semantic alignment and temporal adherence over evaluated streaming baselines, including models retrained on the same data. Ablations and human evaluations validate our design, complemented by spatial conditioning and humanoid execution demonstrations. Our code, data and models will become publicly available.

  5. arXiv · 资产数据集与许可

    Not All Similarity Is Leakage: Distribution-Aware Identity Leakage Assessment in Synthetic Face Datasets

    Synthetic data provides the means for more ethical and less biased facial recognition. Nonetheless, the limitations of generative models and the conditioning process they rely on pose a risk of identity leakage, resulting in synthetic samples that are overly similar to those seen in a generator's training set, mitigating privacy benefits. In this work, we show that previously proposed approaches are affected by two key limitations: high classifier dependence and a tendency to identify leakage with respect to real datasets that were not used in training. To address these shortcomings, we propose a novel metric called Scaled Overlap (SO), which adapts a non-parametric data-copying test for facial recognition and extends it to the identity level. By leveraging a multi-model setup that compensates for errors in individual classifiers, it provides insights into the true strength of identity leakage. Our analysis identifies leaking datasets (CemiFace, SFace2, IDNet) with the most acute leakage introduced by generative model conditioning in CemiFace, demonstrates that real datasets do not contribute equally to generator training, and shows that the majority of synthetic images are drawn from multiple identities rather than relying on a single identity. Ultimately, our study provides new insights into the quality of synthetic datasets and gives more reliable techniques to help the community ensure the trustworthiness of synthetic data.

  6. arXiv · 资产数据集与许可

    FinSafeGuard: A Quality-Driven Synthetic Dataset for Financial AI Safety

    Generating high-quality synthetic datasets at scale remains a fundamental challenge for training specialized AI systems in data-scarce domains. We present a comprehensive methodology that produces 14.3 million quality-filtered synthetic training examples from a structured taxonomy of 632,000 scenario specifications. Our four-stage pipeline ad- dresses critical challenges in synthetic data quality, diversity, and realism through: (1) systematic pre-generation research establishing optimal generation parameters via diminishing returns analysis, (2) conditional auto evolution mechanisms that remediate low-diversity batches during generation, (3) a two-layer semantic deduplication architecture combining lexical and embedding-based methods, and (4) LLM-as-Judge quality evaluation filtering samples based on behavioral correctness and tuple fidelity. Empirical analysis across three domains (Banking, Insurance, Financial Services) demonstrates domain-specific deduplication patterns ranging from 9.9% to 40.6%, with output generation modes exhibiting highest redundancy. Our batch size optimization reveals mode-specific diversity saturation thresholds, enabling principled allocation strategies. The complete pipeline reduces 26.4M raw samples to 14.3M high-confidence examples through merge (20.9% reduction), semantic deduplication (18.7% reduction), and quality-based filtering (15.9% reduction), providing a reproducible framework for domain-specific synthetic data generation with rigorous quality control.

  7. GitHub · vrm-c/UniVRM

    vrm-c/UniVRM 发布 v0.131.4

    Installation You can install UniVRM using the UnityPackage or the UPM Package. The UniVRM supports Unity 2022.3 LTS or later. UnityPackage Download the unitypackage, and drag and drop it to import the package into your project. VRM 1.0: Download VRM 0.x: Download UPM Packages click to expand You can install these UPM packages via Package Manager in UnityEditor. Open the package manager window: Window -> Package Manager Click + at the top-right of the package manager window, then select Add package from git URL... Specify UPM packages using URLs below. [!Warning] Starting with version v0.131.0 , Assets/UniGLTF, VRM, and VRM10 have been moved to the Packages folder. Therefore, please specify path=/Packages/UniGLTF , path=/Packages/VRM , or path=/Packages/VRM10 . UPM package git URL com.vrmc.gltf https://github.com/vrm-c/UniVRM.git?path=/Packages/UniGLTF#v0.131.4 com.vrmc.univrm (VRM 0.x) https://github.com/vrm-c/UniVRM.git?path=/Packages/VRM#v0.131.4 com.vrmc.vrm (VRM 1.0) https://github.com/vrm-c/UniVRM.git?path=/Packages/VRM10#v0.131.4 You can also install via editing Packages/manifest.json directly. // Packages/manifest.json { "dependencies" : { // ... "com.vrmc.gltf" : "https://github.com/vrm-c/UniVRM.git?path=/Packages/UniGLTF#v0.131.4" , "com.vrmc.univrm" : "https://github.com/vrm-c/UniVRM.git?path=/Packages/VRM#v0.131.4" , "com.vrmc.vrm" : "https://github.com/vrm-c/UniVRM.git?path=/Packages/VRM10#v0.131.4" , // ... } } Development Samples click to expand via UnityPackage VRM 1.0 Development Samples: Download VRM 0.x Development Samples: Download via UPM Package You can find Samples in the Package Manager and then submit Import button. Window -> Package Manager -> Packages: In Project -> VRM-1.0 or VRM Release Notes 日本語 English Warning from v0.125.0 . com.vrmc.vrmshaders is removed. com.vrmc.vrmshaders is no longer required in manifest.json. What's Changed [LookAt]モデルの最初の向きを加味する(T-Poseだが違う方向を向いている場合) by @ousttrue in #2839 [VRMC_springBone_limit] 1.0 by @ousttrue in #2840 add nullcheck for normal and uv by @ousttrue in #2841 [unity6.6] fix deprecated by @ousttrue in #2842 [version] UniVRM-0.131.4 by @ousttrue in #2843 Full Changelog : v0.131.3...v0.131.4 官方 Release：v0.131.4。请查看 Release notes、依赖和许可证变化。

  8. arXiv · 创作 Agent 与装配

    EmbodiedWorld-200K: Toward Embodied World Agents via Embodied-Planning Dataset and Interactive World Models

    Building embodied agents that can perceive, reason, and act across diverse 3D worlds is a central research direction toward artificial general intelligence. Yet most existing work remains confined to specific tasks in a single scene and transfers poorly to new environments, leaving general embodied agents for open-world settings under-explored. We refer to such agents as Embodied World Agents (EWAs), whose defining property is the ability to understand diverse visual environments and execute general actions in open-world conditions. To advance EWA research, we study open-world embodied navigation-planning and release EmbodiedWorld-200K, a large-scale dataset of over 200K samples spanning diverse scenes. Each example is equipped with multi-level annotations, ranging from low-level camera-motion trajectories to high-level semantic instructions. The dataset also provides unified evaluation splits and task-specific metrics, supporting offline pretraining and fair evaluation of EWAs. Beyond offline supervision, a defining characteristic of an EWA is its interaction with the environment; training solely on offline data is therefore insufficient, especially for generalizing to scenes not covered in the training distribution. Building dedicated real interactive environments, however, is often prohibitively expensive. To address this, we propose a novel training scheme that employs a video world model as a low-cost environment proxy and performs interactive reinforcement-learning (RL) post-training of the EWA within it. Empirically, pretraining on EmbodiedWorld-200K substantially improves the open-world embodied planning ability of the EWA, and further RL post-training inside a world model yields clear gains on out-of-distribution scenes. The dataset, baseline checkpoints, and training and evaluation code are available at https://xiaokunfeng.github.io/EmbodiedWorld-200K/.

  9. arXiv · 渲染、神经场与导出

    Rethinking Sparse-View 3D Gaussian Splatting: A Few-Shot Learning Perspective on Overfitting Prevention

    Sparse-view novel view synthesis (NVS) is essential for AR/VR and real-time communication applications, yet remains challenging when reconstructing 3D scenes from limited viewpoints. While 3D Gaussian Splatting (3DGS) excels in dense-view scenarios, it suffers from overfitting and geometry degradation under sparse inputs. Current approaches often tackle these problems from a local perspective, addressing either data quality or training regularization individually, lacking a unified and interpretable framework for systematic improvement. We introduce a new perspective that reframes sparse-view 3DGS from a few-shot learning viewpoint, treating each Gaussian primitive as a neuron and leveraging deep learning regularization techniques. Under this unified framework, our approach generates more effective supervision from scarce observations through two strategies: (1) data-related supervision enrichment via optical flow-guided pseudo-view synthesis and joint camera-scene optimization; and (2) training-time regularization via motion-aware dropout and early stopping, which suppress overfitting while preserving dynamic-scene stability. Extensive experiments on the SIGGRAPH Asia 2025 volumetric video challenge sparse-view dataset demonstrate significant improvements over state-of-the-art optimization-based and feed-forward baselines. Our system won first place in the sparse-view track of the SIGGRAPH Asia 2025 challenge.

  10. arXiv · 渲染、神经场与导出

    RxGS: Relightable Gaussian Stylization via BRDF Decomposition

    Reference-based 3D scene stylization transfers the appearance of a reference image onto a reconstructed 3D scene while preserving scene geometry. Recent methods perform this transfer by optimizing the appearance representation of neural scene models including Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS). These methods optimize an RGB appearance representation, in which a single per-point colour combines the effects of material reflectance, environment illumination, together with the shading it produces, and the viewing direction. When stylization modifies this combined colour, the illumination effects present during stylization are baked into the stored appearance, so the stylized scene cannot be rendered under different illumination. To address this limitation, we present RxGS, a relightable 3D scene stylization framework built on a Bidirectional Reflectance Distribution Function (BRDF)-aware 3D Gaussian representation. Rather than modifying a per-point RGB colour, RxGS optimizes only the diffuse albedo field recovered by inverse rendering, while leaving roughness, metallicity, surface normals, and the estimated environment lighting unchanged. As a result, the stylized scene keeps all the components a physically-based renderer needs other than the edited albedo, and can therefore still be relit. RxGS optimizes the albedo field using a nearest-neighbour feature matching loss together with a spatial smoothness prior over Gaussian centres. We evaluate the method using a Geometry Consistency Cost (GCC) based on optical-flow warping error. RxGS achieves the lowest GCC on all evaluated NeRF-Synthetic scenes and remains competitive on real-captured scenes, while additionally supporting relightable stylization, which existing reference-based 3D scene stylization methods do not provide.

  11. arXiv · 渲染、神经场与导出

    SpaRV: Sparse Radiance Voxels for Scalable Geometrically Consistent LiDAR-Visual Mapping and Rendering

    Robotic simulation, navigation, and inspection need photorealistic and geometrically accurate maps. Radiance fields achieve photorealism, but building them from robot data raises three challenges. A robot's trajectory serves its task, not reconstruction, so images from few viewpoints leave the geometry under-constrained. Residual errors in camera poses and intrinsics blur the map. The map's GPU memory grows with scene size, which limits its level of detail in large scenes. We present \textbf{SpaRV}, a \textbf{Spa}rse \textbf{R}adiance \textbf{V}oxel map rendered by differentiable ray marching, which lets irregular LiDAR scans supervise it alongside images. LiDAR depth and occupancy anchor the under-constrained geometry, and normal and smoothness regularizers keep it smooth between LiDAR points. To correct camera errors, SpaRV alternates map training with photometric optimization of the camera poses and focal lengths. To bound GPU memory in training and rendering, SpaRV holds one submap at a time at full resolution and the rest at coarser octree levels. Experiments show that SpaRV keeps the geometry consistent away from the captured trajectory and sharpens textures by optimizing the camera parameters. On one 24\,GB GPU, SpaRV reconstructs a university campus captured in 63{,}582 images and simulates other cameras and LiDARs. Our code will be released at \url{https://github.com/hku-mars/SpaRV}.

  12. GitHub · ronak-create/FableCut

    ronak-create/FableCut 发布 v1.15.0

    Added Agents can look at the footage. fablecut_frame (MCP) and GET /api/frame return a source frame, a contact sheet of evenly spaced frames, or the composed timeline frame at a time, as a JPEG with a text line naming the source, timestamps and clips on screen. text:true returns statistics instead (luma levels on 0–1, dominant colours, texture, saturation and per-cell motion) for agents without vision. Reading a local file by path= requires the X-FableCut-Client header. Media management in the Project bin. Each file's frame rate, codecs, audio channels, sample rate, bitrate and size are read with ffprobe and shown in the bin ( GET /api/media-info ). Search by name, codec, resolution or #tag ; show favorites, used / unused files, a kind, a tag or a collection; sort by any of them or by how often a file is used; a list view with metadata columns. Files can be starred, tagged and grouped into collections. Find timeline instances lists every place a file is used (comps included) and jumps there; Reveal in project goes from a clip to its file. Agents: patch ops updateMedia , setCollection , removeCollection ; the compact view shows metadata, tags and usage. Offline media, relink and replace. A missing file is marked offline in the bin and on the timeline, and its clips show a "Media offline" slate in the monitor and the export. Relink… finds moved files in the project folders by name and size (or matches a chosen folder by name) and keeps every clip's In point; Replace file… puts a new file behind a media entry, trimming clips to fit. Agents: relinkMedia and replaceMedia patch ops, POST /api/relink/candidates ; offline files show in fablecut_status and the compact view. Proxies. Create a light H.264 copy (540p or 720p, frame-accurate, short GOP) of heavy footage from the bin, made by ffmpeg in the background with progress in the bin. With the monitor's Proxy button on (the default), clips and the Source monitor play the proxy while you edit; export, scopes and frame grabs always switch back to the originals. Proxies are a cache in analysis/proxies/ , tied to the source file's version. Agents: fablecut_proxy (make, check, delete); /api/proxy endpoints and SSE event proxy ; the compact view shows proxy:540p . Delivery presets in the Export dialog. YouTube, YouTube Shorts, Instagram Reels, TikTok, X, LinkedIn, Master, Broadcast (ProRes 422 HQ, PCM 24-bit), Web, Archive (lossless FFV1) and Custom, with every setting as a control: resolution and reframing (crop, fit or stretch to another shape), frame rate, codec (H.264, HEVC, AV1, VP9, ProRes, DNxHR, XDCAM HD422, FFV1), encoder (NVENC, Quick Sync and AMF when they work on the machine), CRF / VBR / CBR, max rate, keyframe interval, encoder speed, pixel format, color, container, audio codec / bitrate / rate / channels, a file-name template and metadata. Presets can be saved for every project ( library/delivery/ ). The dialog warns about upscaling, crops, platform length / size / frame-rate limits and estimates the file size. 4:2:2, 4:4:4, 10-bit and lossless outputs get raw RGBA frames. Agents: fablecut_export takes preset and delivery ; setProject {delivery} pins a project default; fablecut_encode_profiles lists the presets and the usable encoders. Encoding profiles still work and sit at the bottom of the list. Batch export and an export queue. A project keeps a list of outputs ( project.outputs ; "+ Typical set" adds master, YouTube 1080p, a 9:16 reel, a square and a 720p preview), and Export checked makes them all with one render per frame rate and range (up to 8 outputs each): each frame is drawn once and every output's ffmpeg encodes it, so four 1080p files take about as long as one. The queue runs outputs back to back with progress, stop, retry and open; one failed output does not stop the rest. An export frame of another shape no longer squeezes an output: a 16:9 file of a project framed for 9:16 reframes the whole canvas. Agents: fablecut_export {outputs:true | [...]} , setProject {outputs} ; REST /api/export/begin {deliveries:[...]} . Loudness normalization and delivery checks. A delivery can normalize the mix to a loudness target (−14 streaming, −16 podcast, −23 EBU R128, −24 ATSC A/85) with a true-peak ceiling, using ffmpeg's two-pass loudnorm ; Broadcast does −23 LUFS / −1 dBTP by default. Every exported file is read back with ffprobe and measured, and checked against what was asked: size, frame rate, frame count, codec, color, audio, length, loudness, true peak and the platform's file-size limit. The report shows in the export queue, after a single export, and in the fablecut_export reply. Changed /media/ serves files in subfolders of media/ too. Chroma key, temperature, tint and RGB split run on the GPU (WebGL2), in preview and export: a 720p keyed frame went from 113 ms to 3.6 ms (RTX 4090) and 123 ms to 12.5 ms (Intel UHD). Output matches the CPU loop within 1–2 levels (a few chroma-key edge pixels up to ~11). Without WebGL2, or on a software rasterizer, the CPU loop still runs. The editor redraws once per project frame , not on every display refresh (240×/s on a 240 Hz panel). While paused it drops to 4 fps after 1.5 s idle. An idle editor sleeps : the loop slows to a 4 fps timer and the audio graph suspends 3 s after the last sound. Idle tab: 83% → 2.2% of a core. 官方 Release：v1.15.0。请查看 Release notes、依赖和许可证变化。

  13. GitHub · scenario-labs/blender-plugin

    scenario-labs/blender-plugin 发布 blender-plugin: v0.12.0

    0.12.0 (2026-10-18) Features blender: apply returned animation clips to the source's rig ( #522 ) ( 8f9349e ) blender: apply saved FBX, OBJ and glTF results in place ( #528 ) ( f8f8b53 ) blender: apply surface edits to several captured objects ( #529 ) ( b40de67 ) blender: match split, reordered and fitted surface results ( #524 ) ( ab19306 ) blender: re-rig a mesh already skinned to an armature ( #520 ) ( 7d7f4cd ) blender: share anonymous usage data, opt-in and off by default ( #617 ) ( 444f1ca ) blender: split parts from the source and fit remeshes in place ( #530 ) ( 4901c7f ) blender: start saved mesh edits from the job's Edit 3D task ( #512 ) ( 4b79229 ) docs: add the Scenario favicon to blender.scenario.com ( #537 ) ( c554583 ) docs: restyle blender.scenario.com and add an in-page search ( #510 ) ( db0119f ) jobs: link uncertain submissions to a user-chosen cloud job ( #527 ) ( 17cbcd5 ) jobs: run result transfers on their own lane with cancel and retry ( #526 ) ( 16c1305 ) mcp: add capture_views, an offscreen multi-view contact sheet ( #597 ) ( 3424fee ) mcp: add Connect buttons that write the MCP client configuration ( #599 ) ( 784be9e ) mcp: connect local agents without a token ( #602 ) ( 9a22761 ) mcp: page scene_summary and bound screenshot sizes for agents ( #562 ) ( 94f3ccd ) mcp: report an oversized execute_python result by its size ( #572 ) ( d2875b1 ) mcp: scale saved model imports to a target size ( #576 ) ( 4c2e1dd ) mcp: start and stop local MCP when Online Access changes ( #574 ) ( c91542e ) schema: store each result's parent within its job ( #610 ) ( 3326b63 ) tools: count extension downloads on blender.scenario.com ( #493 ) ( 7b57141 ) tools: name the plugin on every download event ( #538 ) ( 70fdd02 ) tools: relay opt-in extension usage events and feedback to PostHog ( #616 ) ( 92c26c2 ) tools: tell browsers, scripts and crawlers apart in download counts ( #540 ) ( 701a029 ) ui: ask once to share anonymous usage data after setup ( #631 ) ( 7ee1d24 ) ui: keep feedback text out of Blender's Info editor ( #622 ) ( 883f0ca ) ui: label Feedback in Studio and mention feedback and usage data ( #632 ) ( 14fb642 ), closes #630 ui: offer feedback after errors and after ten generations ( #633 ) ( 1f1ec19 ) ui: send feedback from Blender to Scenario ( #618 ) ( 3638a4f ) Bug Fixes blender: export only the active scene for Edit 3D uploads ( #521 ) ( a1a0a30 ), closes #99 blender: restore the output bit depth after a capture ( #608 ) ( 34f3b0d ) handle malformed MCP requests and GLB, MP4 and audio input ( #553 ) ( 2f094b0 ) jobs: keep delivering results after a transient refresh failure ( #525 ) ( 578353c ) jobs: load a job registry of the wrong JSON shape as empty ( #565 ) ( bf2bf02 ) mcp: answer wait_for_job inside 60-second client limits and accept null options ( #560 ) ( 8981c16 ), closes #558 mcp: bind only the configured MCP port ( #598 ) ( c0e512c ) mcp: list verified local files of every ready result in job_status ( #523 ) ( 6681256 ) mcp: read the selection from the view layer in headless sessions ( #570 ) ( f61eedb ) mcp: resync the view layer before iterating its objects ( #620 ) ( 04c30c3 ), closes #619 Documentation agents: use fix only for behavior users saw ( #534 ) ( 85ad320 ) docs: add agent connection troubleshooting to the user guide ( #561 ) ( 08b7b84 ) docs: check download counts after each release ( #536 ) ( d635975 ) docs: describe delivered in-place edits and point to G14 ( #531 ) ( 3f41109 ), closes #99 docs: keep PR descriptions, commits and doc edits brief ( #542 ) ( 068f182 ) docs: link the OpenSSF Best Practices badge to its final page ( #564 ) ( 78777cc ) docs: merge pull requests through the merge queue ( #603 ) ( af17cee ), closes #587 docs: note how brand assets and inlined scripts are kept ( #539 ) ( 28b116d ) docs: record the 0.11.0 checks that docs still listed as pending ( #517 ) ( e3e59d1 ) docs: record the admin bypass of the merge queue ( #605 ) ( 5d44f43 ), closes #587 docs: record the first approved monthly smoke run ( #586 ) ( ce5a085 ) docs: refresh documentation for 0.12.0 ( #629 ) ( eba1f4c ) docs: show the OpenSSF Best Practices badge in the README ( #551 ) ( a4a53f9 ) docs: state that no Blender job is remotely cancellable today ( #614 ) ( e0513d1 ), closes #494 repo: drop the Scenario logo licence note ( #515 ) ( 901b6ff ) Miscellaneous Chores agents: add Claude Code project settings and local MCP client template ( #518 ) ( 61260f1 ) agents: give Claude Code and Codex the MCP token from .env.local ( #596 ) ( bdf14a5 ) deps: Bump typescript from 5.9.3 to 7.0.2 in the node group across 1 directory ( #548 ) ( c55178e ) deps: keep @types/node on the Node major in .nvmrc ( #547 ) ( 48b743e ) docs: count website visits with Vercel Web Analytics ( #578 ) ( fcee897 ) repo: drop the MIT entry for the Scenario logo ( #583 ) ( a99d2d9 ), closes #29 repo: ignore Vercel CLI project links ( #601 ) ( 2683b69 ) repo: keep PR validation results apart from the checklist ( #544 ) ( e20b662 ) repo: request the maintainer's review on every pull request ( #566 ) ( b5b4de6 ), closes #21 repo: skip the Markdown rewrap in git blame ( #581 ) ( ca4fcf3 ), closes #579 repo: type issues natively instead of with bug and enhancement labels ( #552 ) ( c5358fe ) tools: check that Markdown prose wraps at 100 columns ( #584 ) ( 686d999 ) tools: clear the house-rule failures #519 left on main ( #582 ) ( 2a97310 ), closes #29 tools: deploy the site proxy without a .vercelignore ( #533 ) ( 6a11ddd ) tools: enforce the house rules in check_rules.py, with REUSE.toml and reuse lint ( #519 ) ( 33cf729 ) Continuous Integration ci: add the site proxy tests to the required CI gate ( 7b57141 ) ci: fuzz MCP, upload policy and media parsers with ClusterFuzzLite ( #554 ) ( e2dde2e ) ci: fuzz the site proxy middleware with fast-check property tests ( #549 ) ( 99eb4bd ) ci: install Node tooling with pnpm from a committed lockfile ( #543 ) ( 15c0b1c ) ci: remove the merge queue ( #607 ) ( 553f533 ) ci: run the handbook search tests in the renamed Site job ( db0119f ) ci: run the required checks in the merge queue ( #592 ) ( 3445dcf ) ci: skip GNU.org licence pages that time out in the link check ( #568 ) ( 3169b4f ), closes #567 ci: skip the Blender baseline on documentation-only pull requests ( #535 ) ( a6bc2a7 ) ci: skip the Blender baseline when a pull request only changes the site ( #541 ) ( ff79121 ) ci: title smoke runs with the plan, cap and trigger ( #585 ) ( d1fc2d1 ) deps: Bump astral-sh/setup-uv ( 0f5802c ) deps: Bump astral-sh/setup-uv from 10.2.0 to 10.3.0 in the actions group ( #563 ) ( 0f5802c ) deps: Bump the actions group with 5 updates ( #546 ) ( c3a6444 ) describe the commits check as required, like ci-ok and pr-title ( #577 ) ( b3a5df1 ) mcp: comment the MCP tool definition sizes on pull requests ( #613 ) ( bdb7c00 ) release: attach the provenance bundle to each release ( #550 ) ( 97dcfb1 ) tools: enforce a 5000 CU monthly budget across hosted smoke runs ( #516 ) ( ca59f11 ) Tests blender: check the headless loop's catalog sync by its effect ( #626 ) ( d2ce075 ) blender: give the abandoned-send clock check a margin ( 883f0ca ) blender: run saved mesh-edit tests in the PR baseline ( #511 ) ( 5726f5d ) mcp: budget the context size of local MCP tool definitions ( #595 ) ( 75faf21 ) mcp: expect listed files for saved non-image results in job_status ( #593 ) ( 8e3de55 ), closes #588 mcp: run the non-image job_status test in the PR baseline ( #606 ) ( 396b1d2 ) Styles docs: wrap Markdown prose at 80 columns (mechanical) ( #580 ) ( c2643ef ), closes #579 官方 Release：blender-plugin-v0.12.0。请查看 Release notes、依赖和许可证变化。

10月18日周日
  1. arXiv · 创作 Agent 与装配

    LateMatch: Breaking the Interaction Bottleneck via Cacheable Matching for Efficient Visual Localization

    Visual localization is critical for embodied agents, but deploying state-of-the-art deep feature matchers on edge devices is fundamentally limited by their massive computational cost. Standard deep matchers tightly interleave intra-image context building and inter-image matching at every layer, creating a severe interaction bottleneck. Localizing a single query against K database images forces the system to execute the full network K independent times with no capacity for caching. To break this bottleneck, we advocate for a paradigm shift toward cacheable image matching. We realize this paradigm with LateMatch, a versatile approach that restructures existing matchers by deferring inter-image interaction to their final layers. By isolating intra-image processing into offline-cacheable Context Blocks, only a small set of Interaction Blocks must be executed online. To further optimize LateMatch, we introduce E-SToRE (Entropy-guided Spatial Token Reduction and Expansion). E-SToRE leverages the decreasing Shannon entropy of the late Interaction Blocks to dynamically reduce token counts just prior to cross-image interaction. Evaluated across HPatches, ScanNet, MegaDepth, ETH-3D, and Aachen, LateMatch reduces effective computational cost by over 2x compared to standard deep matchers, while E-SToRE yields an additional 10% to 14% savings with negligible degradation in accuracy.

  2. arXiv · 渲染、神经场与导出

    MCSplat: Multi-View Photometric and Geometric Consistent Feed-Forward Gaussian Splatting for Driving Scenes

    Driving-scene reconstruction via feed-forward Gaussian splatting from sparse multi-view images remains challenging due to two factors. First, cross-camera photometric inconsistency, caused by variations in exposure, ISP, and illumination, can be incorrectly absorbed into Gaussian geometry. Second, unreliable local geometric supervision arises from normal-based constraints near depth discontinuities. To address these issues, we present \textbf{MCSplat}, a multi-view consistent feed-forward Gaussian splatting framework for driving scenes. MCSplat constructs a canonical appearance space from depth-guided multi-camera correspondences to decouple geometry learning from camera-specific appearance. A multi-scale Bilateral Grid Transformer restores raw camera appearance from canonical renderings, enabling photometric supervision without contaminating Gaussian geometry. To improve local geometric reliability, we estimate Fisher-information-based uncertainty from photometric sensitivity and use it to adaptively weight surface-oriented self-supervision. Extensive experiments demonstrate improved reconstruction quality and generalization across driving datasets. Geometry-oriented evaluations further reveal enhanced cross-view depth consistency and local surface-normal agreement, while qualitative results show fewer geometric artifacts and more coherent structures across overlapping views.

  3. arXiv · 资产数据集与许可

    Vision as Proprioception: Estimating Humanoid Robot Pose from Video

    Estimating a humanoid robot's internal state from video has a wide range of real-world applications but remains unexplored. During interactions with a robot, its onboard state information may be unavailable due to privacy, ownership, or communication constraints. Meanwhile, online and AI-generated robot videos provide rich sources of data but typically lack corresponding joint configurations, limiting their direct use in robot learning. To address these issues, we present HumanoidPose, a new framework that treats vision as proprioception to recover full-body joint configurations in world space from video alone. Specifically, we combine pretrained human-pose features with robot kinematics in a unified model that predicts each embodiment's joint configurations. We retarget human motions to sixteen different humanoid embodiments, enabling scalable learning entirely from synthetic data. To improve real-world generalization, we further leverage unlabeled online videos through self-training to develop a 2D robot-pose detector. In our evaluation, HumanoidPose improves joint-angle and joint-position accuracy over the evaluated baselines and reduces joint-angle error by 18-32% relative to the best-performing baseline on each of the three real humanoid platforms. We further show improved global localization and generalization to unseen embodiments. HumanoidPose enables visual understanding of other robots during interaction and extraction of robot motion from online and generated videos.

  4. arXiv · 创作 Agent 与装配

    Compile Once, Run Cheap: Frozen, Portable Skills for Embodied Agents, and What Recursive Self-Improvement Adds

    Embodied AI lacks the data-rich pretraining that gave language models their agents and skills, so new manipulation capability costs demonstrations per task or a frontier model in every episode. We study a third route: a strong general model acts only as an offline author that writes, runs, diagnoses and repairs a manipulation skill in simulation until it passes an exam on layouts drawn after freezing; the frozen skill then runs as code alone or under a cheap caller that reads only its description. We call this skill-level recursive self-improvement Skill-RSI and specify a compile protocol (isolated, audited author; hidden validation layouts; no-feedback and inherit-only controls; tamper-evident ledger) and a reuse protocol that runs the identical files with executors from another model family. On 7 RoboTwin 2.0 tasks with cluttered tables, Claude Code compiled 4 tasks into 7 frozen skills, 5 from a first draft needing only qualification, at USD 4.29-8.22 of author cost each (USD 55 per skill counting the 13 failed campaigns). The other three tasks did not compile in 18 campaigns, including five with GPT-6 and a pre-registered 150-layout run: on the hardest task every campaign scored 78-95 of 100 but had 3-16 safety aborts, failing the gate of 95% success with zero aborts; a 100-episode exam bounds an abort rate only to 3.6%. On hidden layouts, repair that inherits the previous version beat from-scratch rewriting in 8 of 8 controlled campaigns (sign test p = 0.004); an inherit-only lineage attributes most of the gap to inheritance, with development feedback adding a little (3 of 3, p = 0.125). Write-once campaigns found no detectable first-draft effect of practice (minimum detectable effect 14 points). The frozen skills kept 148/150 successes on new layouts with no model and 147/150 with a DeepSeek-flash caller at USD 0.00028-0.00054 per episode; no direct-control baseline was run.

  5. arXiv · 渲染、神经场与导出

    FineSplat: Fine Patches, Local Groups, Detailed Splats

    Feed-forward 3D Gaussian Splatting (3DGS) promises instant scene reconstruction with high-quality novel-view synthesis. Yet, fine-grained geometry and appearance remain difficult to recover, motivating us to revisit how feed-forward models represent, aggregate, and refine detailed visual information. We introduce FineSplat, a feed-forward 3DGS framework built around fine-grained image patches. Since global multi-view attention becomes prohibitively expensive at such high token resolutions, we develop a dual-grouping strategy that combines two complementary forms of locality: view-wise grouping enables efficient interaction among neighboring views, while 3D spatial grouping reconnects geometrically related tokens that may never have interacted under view-wise grouping by lifting them into the 3D scene space. We further leverage reconstruction error as direct feedback, providing an explicit rendering-based signal to recover fine-grained structures and appearance, thereby overcoming the limitations of one-pass feed-forward prediction. Finally, a compact Gaussian generation scheme preserves the benefits of fine-grained features without unnecessarily increasing the number of primitives. Together, these designs enable FineSplat to recover richer geometry and appearance while remaining efficient and compact, achieving state-of-the-art results by a substantial margin, under the same or smaller Gaussian budgets than existing methods.

  6. arXiv · 创作 Agent 与装配

    PRECOCIAL: Pretraining with Code Curriculums via Iterative Search for Accelerated Learning

    Fully synthetic pretraining data generated by symbolic programs can bootstrap downstream training on naturalistic tasks, such as language modeling, but the generating programs have so far been designed by hand. We introduce PRECOCIAL, a method for automatically optimizing programs that generate synthetic pretraining data. Starting from an initial generator program, an LLM iteratively rewrites the program, and each rewrite is scored by pretraining a randomly-initialized transformer on its data and fine-tuning it on the downstream task. Because every score requires training, PRECOCIAL keeps a single network for the entire search, continuing its pretraining for each accepted program instead of training a new network per candidate, making it significantly more efficient. On four language and code datasets, at a fixed fine-tuning budget, optimized programs reach lower perplexity than hand-designed curricula, published procedural generators, random program mutation and the program the LLM writes in the first iteration of the search, and reach the same perplexity as the hand-designed curricula with 9-33% fewer fine-tuning steps. PRECOCIAL also ends with lower perplexity than FunSearch and AlphaEvolve, which train a new network for every candidate, at the same number of evaluations and a lower cost. We also observe that an optimized program generator from one domain transfers to datasets in other domains. Additionally, optimized programs allow for better downstream performance on four different reasoning and problem-solving tasks.

  7. arXiv · 资产数据集与许可

    CasePlay: Self-Play Reinforcement Learning from Case Reports for Medical Reasoning

    Reinforcement learning (RL) is a promising way to improve the clinical reasoning abilities of large language models (LLMs), but current pipelines often rely on expert-labeled questions or costly synthetic-data generation. Meanwhile, the medical literature offers a wealth of high-quality unlabeled documents, including clinical case reports that trace the full clinical course from initial presentation to final treatment decisions. We study whether document-grounded self-play can convert these reports into effective RL training signals. We identify two challenges that limit existing self-play methods in this setting: knowledge-point fixation}, where the proposer repeatedly asks about a narrow set of salient clues in a source report, and difficulty-quality mismatch, where appropriately difficult questions may still be poorly formed or shallow, limiting their value for reasoner training. To address these challenges, we propose CasePlay, a self-play framework with two key components: (a) a Knowledge-Conditioned Proposer, which anchors question generation to case-specific knowledge points so that questions draw on a broader range of evidence from the source report; and (b) a Rubric-Judge Reasoner, which augments answer-accuracy feedback with rubric-based quality feedback on clue integration, reasoning depth, and option quality, guiding the proposer toward reasoning-intensive questions that better support reasoner training. Extensive experiments across seven medically related benchmarks demonstrate that CasePlay outperforms existing self-play baselines and improves the SFT warm-start model by 3.1 points on average, highlighted by a 6.3-point gain on LiveClin. This approach offers an effective path toward continually improving medical reasoning ability through self-play RL.

  8. arXiv · 创作 Agent 与装配

    DialogMem: Evaluating Dialogue-Grounded Embodied Memory in 3D Environments

    Embodied agents in human-centered environments require memory that integrates what they observe with what humans tell them over time. However, existing embodied memories are built primarily from visual experience, leaving open how dialogue with humans can be grounded in the objects encountered throughout that experience. We introduce DialogMem, a two-phase benchmark in which agents build memory from time-aligned streams of posed RGB-D observations and dialogue, and later use it for instance-level object navigation and question answering (QA). DialogMem provides 142 episodes across 27 real-world 3D scenes, with 14,193 QA pairs and 11,890 navigation instructions. We evaluate nine memory representations within a unified framework of memory formation, evidence retrieval, and reasoning for task outputs. In experiments, appending dialogue to visual memory at evaluation time substantially improves QA but yields negligible navigation gains, suggesting that relevant dialogue is present in memory but not grounded to specific objects. Controlled diagnostics show that navigation improves with link quality but falls short of ground-truth object-level memory even under perfect linking. Retrieving the correct object entry and recovering its spatial location remain further bottlenecks. These results identify three bottlenecks for dialogue-grounded embodied memory: grounding and organizing dialogue at object-level entries during formation, retrieving the correct entry, and localizing it accurately.

10月17日周六
  1. arXiv · 资产数据集与许可

    MESS: Multi-Exposure Sequence Synthesis for Generalizable Image Enhancement

    Despite the significant progress in image enhancement under adverse illuminations, such as low-light image enhancement, exposure correction, and backlit image enhancement, existing models show limited generalizability to different scenarios. A major bottleneck lies in the lack of large-scale and diverse training data, as these tasks typically rely on manually captured, precisely aligned image pairs that are expensive to scale up. Although synthetic data have been explored as an alternative, existing synthesis methods are limited in modeling complex adverse illumination conditions and camera imaging pipelines. To address this issue, we propose a novel framework to synthesize realistic training pairs for generalizable image enhancement in the wild. Specifically, we first derive multi-exposure sequences (MES) from high-bit RAW images by rendering them under different exposure settings through an emulated ISP pipeline. Using these RAW-derived MES as supervision, we train a diffusion-based RGB-to-MES generator to synthesize plausible exposure sequences from a single 8-bit RGB image, extending the synthesis pipeline from limited RAW collections to large-scale in-the-wild RGB data. For each source image, we randomly sample one synthesized exposure variant as the degraded input and use the original high-quality RGB image as the target. With the proposed MES Synthesis (MESS) approach, we construct a dataset of 500K training pairs and train a lightweight network, which demonstrates significantly better generalization performance than existing models across various image enhancement tasks. Code and data are available at https://github.com/PolyU-VCLab/MESS.

  2. GitHub · MaxLiu22/Motion2MixamoRig

    MaxLiu22/Motion2MixamoRig 发布 Motion2MixamoRig v0.1.7

    Motion2MixamoRig transfers human motion from a monocular video—or a pose from a still image—onto an Adobe Mixamo-rigged 3D character. This release adds support for Mixamo characters that were edited and re-exported in Cinema 4D. ### Highlights - **Cinema 4D-exported Mixamo FBX support**: use a character you opened and re-exported in Cinema 4D directly as the Mixamo rig, in the Blender add-on or with `--rig`. Details: [`docs/c4d-fbx-support.md`](https://github.com/MaxLiu22/Motion2MixamoRig/blob/v0.1.7/docs/c4d-fbx-support.md). - **Fixed characters falling apart in the GLB**: each mesh piece now keeps its own skin bind matrices. Both were contributed by [@yng47](https://github.com/yng47) in [#8](https://github.com/MaxLiu22/Motion2MixamoRig/issues/8). Thank you! - **User guide in 18 languages**, matching the README. ### What you will notice - In Blender the character now imports as one mesh object per piece (Y Bot: `skinned_mesh_0_Alpha_Surface` and `skinned_mesh_1_Alpha_Joints`) instead of a single `skinned_mesh`. The armature is still named `mixamo`, and the animation is unchanged. - Original Mixamo characters produce the same motion as v0.1.6. ### Install the Blender add-on 1. Download `motion2mixamorig-0.1.7-macos-arm64.zip` from the assets below (do not unzip it). This release has a macOS (Apple Silicon) build only. 2. In Blender 4.2+: **Edit → Preferences → Get Extensions → ∨ → Install from Disk…** 3. Follow the [User Guide](https://github.com/MaxLiu22/Motion2MixamoRig/blob/v0.1.7/docs/user-guide.md). Updating from v0.1.6: remove the old add-on, quit Blender, install this zip, then click **Update Runtime** (the C4D support lives in the runtime code, so this step is required). Downloaded models are kept. ### Important notes * Tested on macOS (Apple Silicon) only. Windows and Linux are untested. * Input currently supports one clearly visible, full-body person. * The SMPL-X body model and Mixamo character must be obtained separately under their respective licenses. * Non-commercial use only, because of upstream model licenses (see `NOTICE.md`). Full changelog: [v0.1.6...v0.1.7](https://github.com/MaxLiu22/Motion2MixamoRig/compare/v0.1.6...v0.1.7)

  3. GitHub · deepseek-ai/deepseek-harness

    deepseek-ai/deepseek-harness 发布 v0.2.1-rc.2

    中文 | English 新增功能 适配 DeepSeek-V4.1-Flash-1017 上线 @LegGasai 适配 DeepSeek Search 与 Fetch API，可在网页搜索设置中开启 @yudshj , @LegGasai New Features Add support for DeepSeek-V4.1-Flash-1017. by @LegGasai Add support for DeepSeek Search and Fetch APIs, which can be enabled in Web Search settings. by @yudshj , @LegGasai Full Changelog: dsh-v0.2.1-rc.1...dsh-v0.2.1-rc.2 官方 Release：dsh-v0.2.1-rc.2。请查看 Release notes、依赖和许可证变化。

  4. GitHub · MaxLiu22/Motion2MixamoRig

    MaxLiu22/Motion2MixamoRig 发布 Motion2MixamoRig v0.1.6

    Motion2MixamoRig transfers human motion from a monocular video—or a pose from a still image—onto an Adobe Mixamo-rigged 3D character. This release makes the Blender add-on usable without a terminal. This plugin has only been tested on macOS so far and hasn't been tested on Windows yet.

  5. arXiv · 角色动作与重定向

    MOMO: Human Motion from Moving Sounds

    Human responses to sound depend not only on its semantic content, but also on the evolving spatial relationship between source and listener, conveyed through spatial audio and its temporal dynamics. Modeling human reactions to such dynamic spatial audio is important for character animation, human behavior analysis, and robotics. As of yet, existing audio-driven motion generation methods primarily focus on non-spatial or static acoustic conditions, leaving this important problem largely underexplored. To bridge this gap, we formulate the novel task of dynamic spatial-audio-driven human motion generation and introduce two motion-captured datasets. We introduce the Dynamic Spatial Audio and Motion dataset (DySAM) as the first dataset and benchmark for this task, providing synchronized binaural audio, 3D human motions, and moving sound source trajectories. Given the importance of dynamic spatial cues of sound sources, we further provide the Dynamic Spatial Audio Localization dataset (DySAL), an auxiliary binaural localization dataset that pairs diverse acoustic events with head poses and moving source trajectories to support spatial audio perception. For benchmarking, we propose MOMO for human Motion generation from Moving Sounds, a diffusion-based framework guided by three coupled principles: localization, understanding, and generation. MOMO performs localization by estimating head-centric spatial direction cues of the moving source directly from raw binaural audio, avoiding reliance on ground-truth source coordinates. It performs understanding by extracting frame-wise audio-content representations, and performs generation by conditioning motion diffusion on both semantic features and dynamic spatial cues. Experiments show that MOMO outperforms existing baselines in motion fidelity, audio-motion alignment, and physical plausibility.

  6. GitHub · deepseek-ai/deepseek-harness

    deepseek-ai/deepseek-harness 发布 v0.2.1-rc.1

    中文 | English 作为 0.2.1 系列的首个候选版本，本版本汇总了自 v0.2.0-rc.2 以来的主要用户和开发者相关变更。 新增功能 允许用户从“自定义插件索引源”中发现和安装插件。 @turtle1999 插件管理页新增「让 Agent 创建插件」入口，可保留草稿进入创造模式；发送需求后才开始执行。 @ZiyaZhang 支持插件启用、停用、卸载及更新后的依赖同步；启用热重载时，更新或重新安装的可替换插件直接生效，无法热重载时提示重启。 @imccyu , @turtle1999 可选插件支持极简模式，Official 插件列表支持按需安装与当前发布版对应的 Claude Code 和 Codex 组合包，并可用于 headless、SDK 和 ACP profile。 @tianyicui , @turtle2099 新增实验性插件的 Session 状态记录接口，供插件写入和读取自己的会话记录。 @tianyicui 新增实验性 Claude Code Mods 兼容层。目前阶段的主要目的是验证 Claude Code Mods API 功能大致为 DeepSeek Harness 插件的一个子集，而非为用户提供实际的完整兼容性。 @tianyicui 新增实验性开发者工具组合包，提供会话原始日志、聊天分组与正文的双向定位，以及内嵌 NodeJS 调试工具；内嵌 DevTools 目前使用英文界面。 @imccyu , @turtle2099 新增实验性思考过程机器翻译插件，可在插件页启用，选择 Bing、Google 或付费 DeepSeek Flash。 @tianyicui , @ZiyaZhang , @imccyu 新增实验性 Git Worktrees 插件，可在插件页启用，由 Agent 创建并进入独立的 Git checkout。 @tianyicui , @turtle2099 自定义模型 API 支持通过UI中配置各模型的推理等级，无需手动修改配置文件。 @LegGasai pi-ai 路由支持按模型能力处理会话中的系统提示词更新，以及工具的动态新增和移除。 @Tuyixiang 极简模式支持读取图片和自动压缩长对话，Web 与桌面端可通过 /compact 指令手动压缩上下文。 @LogosVeritas 全局指令支持读取共享 agents 目录中的 AGENTS.md，可通过 DSH_AGENTS_HOME 指定目录。 @tianyicui 新增 working_directory 工具，并为 TypeScript/Python SDK 提供会话工作目录的读取和切换接口。 @tianyicui SSH 新增独立可执行的 helper 运行时，支持远端文件操作、进程、终端、沙箱和 Node PTC。 @LegGasai Web 版支持在本机 HTTP 访问时登录 DeepSeek 账号。 @lsdsjy 支持识别已发送消息中的密钥和令牌，并默认脱敏按需查看。 @ZiyaZhang 新会话支持预填未发送的提示；恢复草稿或切换工作区时保留文件、目录和会话引用。 @imccyu 新增工作步骤收起时机设置，可选择回答结束后收起，或在下一条普通消息到来时收起。 @imccyu Markdown 文件预览将 YAML 头部元数据显示为清晰的字段列表。 @turtle2099 体验优化 改善自动化任务详情在窄窗口中的布局，关联会话入口会按可用空间收起为图标。 @turtle2099 子智能体可通过配置选用独立 Agent Preset 预设，统一设置角色和工具； @grllll , @Dudu-0223 改进子智能体的委派与等待指引：达到委派层数上限时提前提示，等待子任务期间不再消耗新的目标执行轮次。 @ZiyaZhang , @turtle1999 组合包详情页显示代码来源和当前版本，便于查看安装来源及重装所需的信息。 @turtle1999 插件安装结果显示实际安装版本；因新版本冷却策略安装较旧版本时，提供说明和精确版本安装方式。 @turtle2099 工具调用在准备阶段即可显示已到达的命令说明、文件路径和内容生成进度。 @imccyu Shell 和插件安装详情支持横向查看及复制完整命令；后台 Bash 调用可展开输入和启动回执。 @tianyicui Web 终端更可靠地处理连续输入、粘贴和尺寸调整，并保留进程退出前的最后输出。 @LegGasai PTC Mode 的 Node 执行环境新增可选的 in-memory 模式，多次运行代码可复用变量、函数和已导入模块。 @CreatixChu 改善大量会话的列表加载与 @ 会话搜索体验。 @turtle1999 , @imccyu 优化自动会话标题生成，对同主题的追问、补充和继续请求保留已有合适标题。 @tianyicui , @ZiyaZhang 侧栏 Fork 和子代理目录浏览新增迁移提示，遇到需要迁移的会话时引导用户先打开原会话。 @imccyu 优化侧栏展开规则，让运行中和等待回应的会话更容易找到；会话菜单支持复制会话 ID。 @turtle1999 , @imccyu 支持分别设置界面正文、代码与工具输出、侧栏终端的字体和字号；桌面端可通过 Cmd/Ctrl + 加号或减号统一调整字号。 @turtle2099 , @koalazf99 优化会话和工作区的悬停操作，统一卡片与菜单样式，并改善页头排列及插件页、自动化任务页的窄窗口布局。 @Yifffan , @Kingwl 语音输入新增麦克风选择，可在插件详情页选择输入设备，并查看所选设备的实时输入电平。 @yudshj 允许远程主机访问的 Web 配置使用页面内目录选择器，不再在服务器上弹出原生目录对话框。 @oraluben 优化含超长源码行的文件预览性能，超长行使用无高亮文本展示，普通行继续进行语法高亮。 @turtle2099 Office 文档任务按需加载编辑、渲染和导出指引，减少简单修改及 PDF 导出后的无关检查与预览，并避免在回复中引用已删除的预览图。 @yudshj DeepSeek 账号的模型列表随账号可用模型更新，并隐藏当前客户端版本不支持的模型。 @LegGasai 模型设置更清楚地提示凭据来源和检测状态，支持按需显示输入的 API Key。 @Magolor 问题修复 目标编辑支持多行文本和 Shift+Enter 换行，并改善目标与排队消息编辑器的中文输入及宽度适配。 @turtle2099 修复目标任务停止、恢复并再次停止后，新消息滞留队列的问题。 @turtle2099 修复 Trajectory 将等待、流式输出或重试中的请求标为完成，以及压缩任务提前显示结果的问题。 @turtle2099 , @kermanx 修复后台工作流仍在运行，却因父级工具步骤或对话轮次正常结束而被标为已中断的问题。 @turtle2099 恢复可继续子代理输入栏的图片粘贴和拖放，支持运行中追加图片和完成后发送图片续发消息。 @ZiyaZhang 应用意外退出后，未完成的任务会明确显示“任务已停止”，避免被误认为正常完成。 @turtle2099 修复部分 Bash、PowerShell 和文件修改记录无法展开输出或差异的问题。 @turtle2099 修复凭据、配置和技能文件监听器关闭后的未捕获异常，以及 Agent 关闭时资源提前释放的问题。 @tianyicui , @kermanx 修复部分 Windows 环境下，持久 PowerShell 在大量输出滚屏后迟迟无法结束当前调用的问题。 @Chinesezjc 修复 Windows Desktop 终端退出后，PTY 后台工作进程未随之清理的问题。 @winewill 减少 Linux 下启动命令前的重复检查，降低命令启动开销，并改善较旧 systemd 环境的兼容性。 @imccyu 修复命令执行失败时，未呈现有效的状态信息。 @yudshj 减少新建与恢复多个终端时的重复查询，并避免失效终端在切换面板后反复重试。 @LegGasai 改善 PTC Mode 中缺失的部分指引，并统一 SDK、Web 和 headless 的模型身份说明。 @koalazf99 关闭代码工作视图后仍可选择标准、创造和自定义模式；内置 PTC／极简默认值改为标准，已有会话保持原模式。 @ZiyaZhang 修复启停插件时其他插件的样式被移除、需要刷新页面才能恢复的问题。 @turtle2099 启用开发目录监听后，HMR 可刷新包入口和依赖映射配置，并正确重载仍在运行的原入口。 @imccyu 修复在插件页启用或停用插件后，已打开会话的 / 菜单未更新技能列表的问题。 @grllll 登录请求未收到响应时显示网络检查提示，便于区分连接问题与其他登录失败。 @lsdsjy 修复模型响应流停止输出、且底层传输不响应取消时，请求超过空闲超时仍一直挂起的问题。 @tianyicui 修复插件包元信息不可读，或辅助请求扩展准备失败时，整个 DeepSeek 请求被阻断的问题。 @turtle1999 , @turtle2099 修复模型设置刷新期间仍可基于旧数据编辑或删除配置，导致刚保存的选择未出现在编辑器中的问题。 @tianyicui 凭据文件中出现 DSH 不使用的额外字段时，不再阻止应用启动。 @turtle2099 桌面端默认使用系统分配的端口，避免 Windows 保留端口范围导致应用无法启动。 @turtle2099 修复展开右侧栏后，会话标题栏中的后台任务列表被侧栏遮挡或裁切的问题。 @turtle2099 修复 Windows Desktop 图片预览遮罩及关闭按钮与标题栏重叠、侧栏按钮悬停圆角不一致的问题。 @turtle2099 修复 Windows 安装器对磁盘根目录的校验，拒绝带重复末尾路径分隔符的根目录写法。 @tianyicui 修复 macOS 应用的小尺寸 Finder 图标，以及复制和替换应用提示中的图标损坏问题。 @07akioni 修复 Windows 和 Linux 下右侧栏收起时，在输入框按 Page Up 导致整个界面向左偏移的问题。 @imccyu 修复侧栏列表重排或重新显示后，会话加载动画不同步的问题。 @turtle2099 修复 /goal、/plan 等命令与参数交界处的输入异常及草稿与显示不一致，支持整体选中、删除命令，并在撤销或刷新页面后恢复。 @LegGasai 修复切换会话、收起工作过程或发送后续消息时，对话区域出现大块空白、滚动跳动的问题。 @imccyu 修复重启应用后部分会话无法继续，以及切换到 DeepSeek 后因历史消息兼容问题导致请求失败的情况。 @imccyu , @yudshj 修复 macOS 添加工作区时目录选择器无法粘贴目录路径，以及取消选择仍添加工作区的问题。 @turtle2099 修复原预设被删除或调整后，恢复空白会话、切换工作区可能失败的问题。 @LegGasai 修复卸载会话存储插件时，最后的会话记录可能丢失的问题。 @turtle1999 修复本地 SenseVoice 语音识别开启 VAD 后，录音开头被截断、首字或音节丢失的问题。 @yudshj 修复远程 Web 在网络拥塞时被误判为断线，以及刷新后语言、主题等偏好丢失的问题。 @turtle2099 , @oraluben 其他变更 API Key 读取优先级变更：Web端/桌面端优先采用已保存的凭据；headless 默认改用环境变量和 .env，不读取已保存的密钥。 @Magolor 移除运行时 invariant 插件及各包的 ./invariant 导出；使用这些诊断入口的开发者需相应调整扩展或自定义 profile。 @turtle2099 子路径插件不再读取独立的 package.json ，显示文本和图标须通过对应子路径导出；插件开发模板同步补齐本地化信息与图标。 @turtle1999 第三方插件的版本兼容性优先按 engines.dsh 判断；未声明时沿用原有规则。插件作者需检查声明是否覆盖支持的 DSH 版本。 @turtle1999 输入区统计扩展拆为 activity 和 usage 两个独立入口；覆盖旧 stats 整行的插件需要更新注册 ID。 @turtle1999 为用户消息、文本与图片展示、新会话页及模型选择器添加扩展位，支持插件定制对应内容。 @imccyu , @LegGasai 自动化任务改为 Web 内置能力，提醒工具按模式提供：标准、创造和 PTC 模式可用，极简模式和子代理不可用；旧实验组合包选择会自动清理，已有任务保留。 @Chinesezjc , @turtle1999 关闭计划审阅以补充修改意见时，直接结束当前轮次并等待用户反馈，避免额外调用模型。 @turtle2099 统一子代理执行与完成通知，工具立即返回 child ID；自定义 SDK runtime 需支持 session/wait。 @Dudu-0223 Agent Team 消息改为直接投递目标 Inbox，移除独立 outbox、自动重试和重发去重；相关配置与结果字段需迁移。 @Dudu-0223 Python PTC 接入 Session 文件沙箱策略；自定义组合需提供 sandbox 和 sandboxPolicy 服务。 @tianyicui 移除工具展示的 both 混合模式，仅保留 native 和 ptc；使用旧模式的配置与 API 调用需要更新。 @tianyicui 移除 agent-instructions 插件逐行 dshHome 配置，指令目录改从进程的 DSH_HOME 或默认目录读取。 @tianyicui 默认 SDK profile 从固定的编程 Agent 改为通用 AI Agent 身份，适用于一般任务。 @Dudu-0223 使用 --from-default-profile 重建已有自定义配置时，会替换该配置及其插件和设置。 @turtle2099 TypeScript 和 Python SDK 支持 resume 功能，恢复已保存的会话，继续使用原有会话 ID 和工作目录。 @koalazf99 Python SDK 的 Office 引擎和完整创作环境改为按需下载。 @yudshj 完善网关授权与连接机制，支持按调用方身份控制资源访问，并独立管理 API 连接及其交互参与方式。 @imccyu Web 支持通过 --public-url 指定显示、打开及提供给模型的对外访问地址，支持带路径前缀的反向代理入口。 @oraluben Web 支持绑定指定的 IPv4/IPv6 地址，并通过 --tls-cert 和 --tls-key 配置原生 HTTPS 监听。 @oraluben pi-ai 更新至 1.0.2，包含上游模型目录更新和适配器修复。 @tianyicui 更新第三方模型支持；Azure 提供方名称统一为 azure，使用旧名称 azure-openai-responses 的提供方配置需迁移。 @LegGasai 更新运行时工具、MCP 协议、图片处理和客户端渲染依赖，纳入兼容版本的安全修复。 @tianyicui This first release candidate for 0.2.1 summarizes the major user- and developer-facing changes since v0.2.0-rc.2. New Features Discover and install plugins from custom plugin index sources. by @turtle1999 Add a “Let Agent create a plugin” entry to plugin management. It opens Creator mode while preserving the draft and starts work only after the request is sent. by @ZiyaZhang Support dependency synchronization when plugins are enabled, disabled, uninstalled, or updated. With hot reload enabled, replaceable plugins take effect directly after an update or reinstallation; changes that cannot be hot-reloaded request a restart. by @imccyu , @turtle1999 Support optional plugins in Minimal mode. The Official plugin catalog offers on-demand installation of Claude Code and Codex bundles matching the current release, for use in headless, SDK, and ACP profiles. by @tianyicui , @turtle2099 Add Session state-record APIs for experimental plugins to write and read their own session records. by @tianyicui Add an experimental Claude Code Mods compatibility layer. At this stage, the main goal is to verify that the Claude Code Mods API's capabilities are broadly a subset of what DeepSeek Harness plugins can do, rather than to offer users complete, practical compatibility. by @tianyicui Add an experimental Developer Tools bundle with raw session logs, bidirectional navigation between chat groups and the conversation, and embedded NodeJS debugging tools. The embedded DevTools currently uses an English interface. by @imccyu , @turtle2099 Add an experimental reasoning translation plugin, enabled from the Plugins page with Bing, Google, or paid DeepSeek Flash as the provider. by @tianyicui , @ZiyaZhang , @imccyu Add an experimental Git Worktrees plugin that can be enabled on the Plugins page for Agents to create and enter separate Git checkouts. by @tianyicui , @turtle2099 Configure reasoning levels for individual custom API models in the UI without editing configuration files manually. by @LegGasai Support in-conversation system prompt updates and dynamic tool additions and removals on pi-ai routes according to model capabilities. by @Tuyixiang Minimal mode can read images and automatically compact long conversations. Web and Desktop users can run the /compact command to compact context manually. by @LogosVeritas Load global instructions from AGENTS.md in the shared agents directory, with DSH_AGENTS_HOME available to select its location. by @tianyicui Add the working_directory tool and TypeScript/Python SDK APIs for reading and changing a session’s working directory. by @tianyicui Add standalone executable SSH helper runtimes supporting remote file operations, processes, terminals, sandboxing, and Node PTC. by @LegGasai The Web app supports DeepSeek account sign-in when accessed over local HTTP. by @lsdsjy Detect keys and tokens in sent messages and mask them by default, with an option to reveal them. by @ZiyaZhang Support unsent initial prompts for new sessions and preserve file, folder, and session references when restoring drafts or moving them between workspaces. by @imccyu Add a setting to collapse work details either when the response finishes or when the next regular message arrives. by @imccyu Display YAML frontmatter as a readable field list in Markdown file previews. by @turtle2099 Improvements Improve Automation task details in narrow windows by collapsing the linked-session action to an icon when space is limited. by @turtle2099 Configure spawned subagents with independent Agent Presets to manage their roles and tools together. by @grllll , @Dudu-0223 Improve subagent delegation and waiting guidance: announce delegation depth limits in advance and wait for running subtasks before starting another goal round. by @ZiyaZhang , @turtle1999 Show a bundle’s code source and current version in its details, making installation and reinstallation information easier to find. by @turtle1999 Show the version actually installed and explain how to request an exact version when pnpm’s release-age policy selects an older release. by @turtle2099 Show available command descriptions, file paths, and content-generation progress while tool calls are still preparing. by @imccyu Support scrolling and copying full commands in Shell and plugin installation details, and expand background Bash inputs and launch receipts. by @tianyicui Handle continuous input, pastes, and resizing more reliably in the Web terminal, and preserve the final output when a process exits. by @LegGasai The Node runtime for PTC Mode adds an optional in-memory mode that reuses variables, functions, and imported modules across code executions. by @CreatixChu Improve large session list loading and @ session search. by @turtle1999 , @imccyu Improve automatic session titles by retaining adequate titles for same-topic follow-ups, additional details, and requests to continue. by @tianyicui , @ZiyaZhang Add migration guidance to sidebar forks and subagent catalog browsing, directing users to open the original session when migration is required. by @imccyu Improve sidebar expansion so running sessions and sessions awaiting a response are easier to find. Session menus support copying the session ID. by @turtle1999 , @imccyu Choose fonts and sizes separately for interface text, code and tool output, and sidebar terminals. On Desktop, use Cmd/Ctrl + plus or minus to adjust font sizes together. by @turtle2099 , @koalazf99 Refine session and workspace hover interactions, unify card and menu styling, and improve header ordering and narrow-window layouts on plugin and automation pages. b

  7. GitHub · mutuals-inc/character-harness

    mutuals-inc/character-harness 发布 v0.1.0 — macOS Apple Silicon preview

    macOS Apple Silicon player v0.1.0 (unsigned). Download character-player-0.1.0-macos-arm64.dmg , open it and drag Character Player to Applications. macOS blocks it the first time: open it once, then go to System Settings > Privacy & Security, click Open Anyway and confirm (on macOS 14 and earlier, Control-click the app and choose Open). The compiled engine and FoxGlove model and animations are included; no Kaigen installation or licence is required. Run the full harness Install Node.js 23.6 or later and download this tag's Source code (zip) . Extract the source, open Terminal in its root (beside package.json), and run npm ci --omit=dev . Copy .env.example to .env and add your own XAI_API_KEY and ELEVENLABS_API_KEY . Set HARNESS_PLAYER in .env to the installed player: HARNESS_PLAYER = " '/Applications/Character Player.app/Contents/MacOS/'* " Run npm run check -- characters/foxglove , then npm start from the harness root. The runtime opens the player automatically. Allow the microphone when macOS asks and speak first . The player alone does not run the conversation; Node and provider API keys are required. FoxGlove needs no additional animation setup. Provider usage is billed to your accounts. For startup problems, read the Terminal output and .cache/player.log . See the harness quickstart , macOS installation guide , custom character guide , and architecture . Verify the download with shasum -a 256 -c --ignore-missing SHA256SUMS-macos-arm64 . Built from v0.1.0-macos-arm64 ( 8bd1f91 ), Kaigen 0.5.8.003-internal. SHA-256: 2f3c872abd16fb9270bd85f80957a63c056ea32359108801f63bc9dce8166243 character-player-0.1.0-macos-arm64.dmg e227594f807186d612ce89d4be9faa96c2e505de9d1b87a9b2e64573acab06a0 character-player-0.1.0-macos-arm64.zip 官方 Release：v0.1.0-macos-arm64。请查看 Release notes、依赖和许可证变化。

  8. arXiv · 资产数据集与许可

    Landmark-Free and Scale-Blind Multi-View Registration of Preoperative Liver Models to Intraoperative Surfaces

    Augmented-reality guidance in laparoscopic liver resection requires aligning preoperative 3D liver models from CT/MRI with intraoperative surfaces. Monocular reconstruction in most cases cannot guarantee scale, albeit rigid-motion registration requires scale as an input. Our proposed framework ``SimCompose'' estimates the full similarity transformation $\Sim$, including scale, without relying on landmarks and applies the same transformation across multiple views with known relative geometry. However, reliable detection of registration failures remains challenging in surgical settings, where ground-truth alignment is generally unavailable. For this, we propose a geometry-based selector that evaluates candidate transformations without requiring ground truth. Large disagreement among these transformations is used to flag potentially unreliable registrations. Finally, we propose to use a non-rigid refinement based on a low-dimensional deformation model from Neo-Hookean simulations in the Simulation Open Framework Architecture (SOFA), without additional network training. SimCompose is trained exclusively on synthetic partial views. Its encoder processes point clouds normalised to a unit root-mean-square radius, without access to metric scale, and filters each view before multi-view fusion. Evaluation on unseen synthetic data (300 single-view pairs per run; 100 three-view registrations) shows that multi-view integration and candidate selection increase the success rate under stringent $\Sim$ accuracy criteria from 38.4% for a single-view, single-method configuration to 77%. Existing learning-based baselines do not satisfy these criteria even when provided with the exact scale. On five in-vivo patients with 92 keyframes, the rigid registration stage reduces surface error more than three-fold relative to learning-based baselines with external scale information.

  9. arXiv · 创作 Agent 与装配

    Self-Improving Video-to-Action Models

    Video generative models are an emerging tool for embodied agents: they can serve as scalable visual planners, generating videos that depict how to complete previously unseen tasks. However, executing these plans typically requires a video-to-action or "inverse dynamics" model trained on large, human-collected video-action datasets, severely limiting scalability. We introduce Mirror, a self-supervised approach that learns video-to-action mappings through environment interactions, without prior action data or hand-crafted rewards. Mirror learns by attempting to follow a visual plan by executing its best guess of the underlying actions in the environment. Mirror then trains on the resulting video-action pairs, incrementally self-improving and following the visual planner more closely. Across Atari, OGBench, and DeepMind Control, Mirror outperforms leading learning-from-observation methods with learned video plans, and achieves near-expert performance with oracle plans. On Minecraft BASALT, it outperforms methods with access to expert actions or hand-engineered task knowledge. In Minecraft survival, it is the first method to acquire an iron pickaxe through self-improvement alone, without rewards or expert action labels. In the offline setting, with expert action labels but no environment interaction, it sets a new state of the art in obtaining diamonds. Our results suggest that advances in video models can translate into increasingly capable agents without a corresponding need for hand-crafted rewards or large collections of expert action data.

  11. GitHub · mutuals-inc/character-harness

    mutuals-inc/character-harness 发布 v0.1.0 — Windows x64 preview

    Windows x64 preview of Mutuals Character Harness v0.1.0 . Downloads character-player-0.1.0-windows-x64.zip — the native player, compiled engine, FoxGlove model, animations and runtime assets. Extract the whole ZIP and keep its files together. The executable cannot run alone. SHA256SUMS-windows-x64 — SHA-256 checksum for that ZIP. Matching harness source ZIP — the Node conversation runtime and character configuration. This is also GitHub's “Source code (zip)” download below. No Kaigen SDK, engine licence, compiler or Developer Mode is needed to run this package. The player handles rendering, microphone capture and playback; the Node runtime runs the conversation. Install both using the steps below. FoxGlove's model and animations are already included: no animation setup is needed for the default character. Run the whole harness on Windows Install Node.js 24 or later from nodejs.org . Open a new PowerShell window and check: node -- version npm.cmd -- version Download the player ZIP and checksum above. From your download folder, run: Get-FileHash .\character - player - 0.1 . 0 - windows - x64.zip - Algorithm SHA256 Get-Content .\SHA256SUMS - windows - x64 The hashes must match (letter case does not matter). Extract the ZIP to a permanent location such as C:\Users\you\Apps\ . You should have a folder containing Character Player.exe and cooked\ ; keep the entire folder together. This preview is unsigned, so Windows may display a publisher warning. Download and extract the matching harness source ZIP above into a separate folder. In PowerShell, change into the extracted harness root: the folder containing package.json , .env.example , characters\ and runtime\ . Run: npm.cmd ci -- omit = dev Copy-Item .env.example .env notepad .env For an existing setup, keep your .env instead of overwriting it. In .env , enter your own XAI_API_KEY and ELEVENLABS_API_KEY . They provide FoxGlove's chat model, speech recognition and voice. The keys need access to the models/voice configured in characters/foxglove/character.yaml ; provider usage is billed to your accounts. Also set the actual path to the extracted player, keeping both quote levels for paths with spaces: HARNESS_PLAYER = ' "C:/Users/you/Apps/character-player-0.1.0-windows-x64/Character Player.exe" ' Replace you and the folder path with yours, then save .env . Do not publish or share your keys. From that same harness root, validate the character configuration, then start it: npm.cmd run check -- characters / foxglove npm.cmd start The check validates configuration; it does not test your provider accounts. start runs the Node runtime and opens the player automatically using HARNESS_PLAYER . Keep PowerShell open. In Windows Settings → Privacy & security → Microphone, enable microphone access and access for desktop apps. Speak first : FoxGlove waits for you before replying. Closing the player stops the runtime; Ctrl+C in PowerShell also stops the session. Help and customization Harness quickstart — runtime setup and API keys. Windows player installation — player paths, package layout and microphone access. For startup failures, read the PowerShell output and .cache/player.log under the harness source folder. Bring your own character — your own VRM model and animations. Then run npm.cmd start -- characters/<id> . Character manifest — configure models, voice, rig and extensions. Architecture and your first extension — understand and extend the conversation loop. Version and provenance This release is Windows x64 only , tagged v0.1.0-windows-x64 . macOS and Linux builds can use their own platform tags for the same app version when available; this release contains neither. The existing player ZIP is unchanged. Built from commit e73eca6b242b4f29fc4e9f9f0f78db9972273a25 with Kaigen 0.5.8.003-internal . Project code and Mutuals-owned FoxGlove assets use Apache-2.0; included third-party work retains its own licences. 官方 Release：v0.1.0-windows-x64。请查看 Release notes、依赖和许可证变化。

10月16日周五
  1. GitHub · deepseek-ai/deepseek-harness

    deepseek-ai/deepseek-harness 发布 v0.2.1-alpha.5

    中文 | English 问题修复 修复在插件页启用或停用插件后，已打开会话的 / 菜单未更新技能列表的问题。 @grllll 减少新建与恢复多个终端时的重复查询，并避免失效终端在切换面板后反复重试。 @LegGasai 改善 PTC Mode 中缺失的部分指引，并统一 SDK、Web 和 headless 的模型身份说明。 @koalazf99 修复 Windows 和 Linux 下右侧栏收起时，在输入框按 Page Up 导致整个界面向左偏移的问题。 @imccyu 其他变更 新增 conversation.input.model.label Slot，支持插件按模型渠道自定义输入框模型选择器中的名称展示。 @LegGasai Bug Fixes Fix the skill list in an open Session’s / menu not updating after a plugin is enabled or disabled on the Plugins page. by @grllll Reduce redundant queries when creating or restoring multiple terminals, and prevent failed terminals from retrying repeatedly when switching panels. by @LegGasai Improve incomplete guidance in PTC Mode, and align model identity prompts across SDK, Web, and headless profiles. by @koalazf99 Fix Page Up in a text input shifting the entire interface to the left on Windows and Linux when the right sidebar is collapsed. by @imccyu Chores Add the conversation.input.model.label Slot so plugins can customize model labels by provider in the composer’s model selector. by @LegGasai Full Changelog: dsh-v0.2.1-alpha.4...dsh-v0.2.1-alpha.5 官方 Release：dsh-v0.2.1-alpha.5。请查看 Release notes、依赖和许可证变化。