Embodied AI Daily具身智能科研日报

具身智能早报

关注机器人研究、开源与产业的新进展

‹ ›
具身智能日报 · 2026-11-10
本期导读

本期整理机器人学习、系统与工具的研究进展;各条附来源依据与阅读边界。

7 条内容 · 3 篇重点解读 · 4 条简讯

On Agentic Manipulation Policies:Figure 1 : Zero-shot real-world manipulation with AMP1. AMP1 solves diverse, long-horizon manipulation tasks directly in the real world from a natural language instruction without task-specific fine-tuning or predefined skills. It devises complex strategies, such as ( A ) searching for hidden objects, ( B ) using a balance scale as a measuring instrument, ( C ) planning and executing a winning strategy in a game of chess, ( D ) combining non-prehensile and prehen
Figure 1 : Zero-shot real-world manipulation with AMP1. AMP1 solves diverse, long-horizon manipulation tasks directly in the real world from a natural language instruction without task-specific fine-tuning or predefined skills. It devises complex strategies, such as ( A ) searching for hidden objects, ( B ) using a balance scale as a measuring instrument, ( C ) planning and executing a winning str图源:论文作者 / arXiv
arXiv cs.CV filtered精读

On Agentic Manipulation Policies

Chen Chen、Omid Bagherieh 等 · 13 位作者

The authors present AMP1, a reference harness in which a generalist AI model commands the manipulator directly and adapts its actions as the task unfolds, running unchanged across single-arm, bimanual, and mobile manipulation. On the RoboLab-120 benchmark, AMP1 reaches an 85.6% success rate zero-shot with no demonstrations or task-specific training, compared to 49.2% and 42.9% for leading VLA and WAM policies. In real-world trials spanning twelve tasks and three robot instances, the agent succeeds zero-shot in 104 of 108 trials, reducing failure rates by 20x and 15x relative to strong VLA and WAM baselines. The basic harness supplies no grasping, pushing, or placing skills; the agent develops its own strategies, such as searching for hidden objects, using objects as tools, and calibrating its cameras. With the addition of a persistent memory store, an augmented harness lowers the failure rate on RoboDojo bimanual manipulation tasks by 4.6x relative to zero-shot deployment.

机器人基础模型Agentic Manipulation PoliciesZero-shot Learning
论文
展开解读与证据

The authors present AMP1, a reference harness in which a generalist AI model commands the manipulator directly and adapts its actions as the task unfolds, running unchanged across single-arm, bimanual, and mobile manipulation. On the RoboLab-120 benchmark, AMP1 reaches an 85.6% success rate zero-shot with no demonstrations or task-specific training, compared to 49.2% and 42.9% for leading VLA and WAM policies. In real-world trials spanning twelve tasks and three robot instances, the agent succeeds zero-shot in 104 of 108 trials, reducing failure rates by 20x and 15x relative to strong VLA and WAM baselines. The basic harness supplies no grasping, pushing, or placing skills; the agent develops its own strategies, such as searching for hidden objects, using objects as tools, and calibrating its cameras. With the addition of a persistent memory store, an augmented harness lowers the failure rate on RoboDojo bimanual manipulation tasks by 4.6x relative to zero-shot deployment.

为什么值得关注

The analysis suggests that agentic manipulation policies are substantially more general and robust than specialist manipulation policies, providing a promising foundation for robot manipulation.

研究问题

Recent work on robot manipulation has focused on specialist policies trained on demonstrations, such as vision-language-action models (VLAs) and world-action models (WAMs).

方法原理

The authors present AMP1, a reference harness in which a generalist AI model commands the manipulator directly and adapts its actions as the task unfolds, running unchanged across single-arm, bimanual, and mobile manipulation.

结果与证据

On the RoboLab-120 benchmark, AMP1 reaches an 85.6% success rate zero-shot with no demonstrations or task-specific training, compared to 49.2% and 42.9% for leading VLA and WAM policies. In real-world trials spanning twelve tasks and three robot instances, the agent succeeds zero-shot in 104 of 108 trials, reducing failure rates by 20x and 15x relative to strong VLA and WAM baselines. The basic harness supplies no grasping, pushing, or placing skills; the agent develops its own strategies, such as searching for hidden objects, using objects as tools, and calibrating its cameras. With the addition of a persistent memory store, an augmented harness lowers the failure rate on RoboDojo bimanual manipulation tasks by 4.6x relative to zero-shot deployment.

阅读边界
依据本次返回的正文节选整理,未核验未返回的章节、附录或独立复现。

作者:Chen Chen、Omid Bagherieh、Naveen Nikamanth Appiah Balaji、Jesse T. Gonzalez、Negin Heravi、Fahad Islam、Xueting Li、Jean-Baptiste Regli、Kejia Ren、Petr Švec、David J. Yoon、Matthias Müller、Vladlen Koltun

追问 / 纠错 / 阅读偏好 · 交给 Agent

评论 · —

昵称自填。评论会展示给其他读者,请勿填写私人信息。

正在读取评论…

Computer-Use Agents Are Embodied Generalists:Figure 1: Computer-use agents can operate any software through keyboard and mouse actions while observing the screen. We argue that physical environments are no different. Once exposed through software, they become just another interface for the same agent.
Figure 1: Computer-use agents can operate any software through keyboard and mouse actions while observing the screen. We argue that physical environments are no different. Once exposed through software, they become just another interface for the same agent.图源:论文作者 / arXiv
arXiv cs.LG filtered精读

Computer-Use Agents Are Embodied Generalists

Pranjal Aggarwal、Sean Welleck

The authors formulate embodied tasks as interaction with digital software that renders an agent's view of a physical environment on a screen and provides controls for acting in it. A computer-use agent (CUA) performs embodied tasks through the same screenshots and keyboard and mouse actions that it uses to operate ordinary software. Using a single CUA design and changing only the software, entries won first-, second-, third-, and sixth-place prizes across four ECCV workshop competitions without competition-specific agent tuning.

机器人基础模型Computer-Use AgentsEmbodied Generalists
论文
展开解读与证据

The authors formulate embodied tasks as interaction with digital software that renders an agent's view of a physical environment on a screen and provides controls for acting in it. A computer-use agent (CUA) performs embodied tasks through the same screenshots and keyboard and mouse actions that it uses to operate ordinary software. Using a single CUA design and changing only the software, entries won first-, second-, third-, and sixth-place prizes across four ECCV workshop competitions without competition-specific agent tuning.

为什么值得关注

The results suggest that skills CUAs learn on software, such as visual grounding and precise clicking, carry over to embodied tasks, advocating for computer-use agents as embodied generalists.

研究问题

Embodied AI is poised for a shift where general-purpose models replace handcrafted methods, similar to trends in other areas of AI.

方法原理

The authors formulate embodied tasks as interaction with digital software that renders an agent's view of a physical environment on a screen and provides controls for acting in it. A computer-use agent (CUA) performs embodied tasks through the same screenshots and keyboard and mouse actions that it uses to operate ordinary software.

结果与证据

Using a single CUA design and changing only the software, entries won first-, second-, third-, and sixth-place prizes across four ECCV workshop competitions without competition-specific agent tuning.

阅读边界
依据本次返回的正文节选整理,未核验未返回的章节、附录或独立复现。

作者:Pranjal Aggarwal、Sean Welleck

追问 / 纠错 / 阅读偏好 · 交给 Agent

评论 · —

昵称自填。评论会展示给其他读者,请勿填写私人信息。

正在读取评论…

Seeing What Matters: Benchmarking Task-Conditioned Actionable Spatial Belief for Manipulation:Figure 1 : Seven-model evidence diagnostics. Left: success versus target-cue rate. Right: success versus wrong-command cue rate. Seven strict-clean VLMs each have 1,500 evaluation records, preview twelve RGB candidates, and select exactly three views. Dots show point estimates; icons label models. Cue rates are separate evaluator diagnostics, not complementary fractions.
Figure 1 : Seven-model evidence diagnostics. Left: success versus target-cue rate. Right: success versus wrong-command cue rate. Seven strict-clean VLMs each have 1,500 evaluation records, preview twelve RGB candidates, and select exactly three views. Dots show point estimates; icons label models. Cue rates are separate evaluator diagnostics, not complementary fractions.图源:论文作者 / arXiv
arXiv cs.CV filtered精读

Seeing What Matters: Benchmarking Task-Conditioned Actionable Spatial Belief for Manipulation

Yina Jian、Mingkun Xu

The authors introduce Seeing What Matters, a benchmark for task-conditioned actionable spatial belief in simulated tabletop manipulation across coffee, drawer, and microwave domains. The benchmark provides 860 frozen evaluation cases and measures whether a view-selection policy acquires the observations its command actually needs under closed-loop and functional-OOD protocols. A diagnostic reference pipeline reaches 98.8% closed-loop and 98.9% functional-OOD success, while seven vision-language models evaluated on the same cases reach only 26-43%.

学习与仿真Spatial BeliefBenchmark
论文
展开解读与证据

The authors introduce Seeing What Matters, a benchmark for task-conditioned actionable spatial belief in simulated tabletop manipulation across coffee, drawer, and microwave domains. The benchmark provides 860 frozen evaluation cases and measures whether a view-selection policy acquires the observations its command actually needs under closed-loop and functional-OOD protocols. A diagnostic reference pipeline reaches 98.8% closed-loop and 98.9% functional-OOD success, while seven vision-language models evaluated on the same cases reach only 26-43%.

为什么值得关注

The performance gaps suggest that spatial belief for manipulation should be evaluated by whether it acquires the evidence the current command requires, not by scene-level completeness.

研究问题

Spatial-reasoning benchmarks evaluate whether a model can describe a scene, but manipulation requires acquiring the particular evidence that makes the current command executable.

方法原理

The authors introduce Seeing What Matters, a benchmark for task-conditioned actionable spatial belief in simulated tabletop manipulation across coffee, drawer, and microwave domains. The benchmark provides 860 frozen evaluation cases and measures whether a view-selection policy acquires the observations its command actually needs under closed-loop and functional-OOD protocols.

结果与证据

A diagnostic reference pipeline reaches 98.8% closed-loop and 98.9% functional-OOD success, while seven vision-language models evaluated on the same cases reach only 26-43%.

阅读边界
依据本次返回的正文节选整理,未核验未返回的章节、附录或独立复现。

作者:Yina Jian、Mingkun Xu

追问 / 纠错 / 阅读偏好 · 交给 Agent

评论 · —

昵称自填。评论会展示给其他读者,请勿填写私人信息。

正在读取评论…

HiMEM introduces a hippocampus-inspired memory consolidation mechanism that synthesizes episodic experiences into compact semantic memories to derive generalizable principles. The framework includes an episodic memory pruning mechanism to remove redundant memories and a flexible memory composition mechanism to leverage multiple memories for cross-scenario generalization. Experiments on LIBERO demonstrate that HiMEM improves the average success rate of pi_0 from 91.88% to 98.20%.

更多

依据摘要或发行说明整理,未核验完整正文或独立复现。

追问 / 纠错 / 阅读偏好 · 交给 Agent

评论 · —

昵称自填。评论会展示给其他读者,请勿填写私人信息。

正在读取评论…

InstantHMR is a family of open-source single-image regressors based on the Momentum Human Rig parametric model, suited for mobile inference. The approach recovers a 6,890-vertex body mesh in 5.5 ms per person on a Samsung Galaxy S23 NPU and 0.98 ms on an NVIDIA RTX 4070 GPU. It achieves competitive accuracy with 40.0 mm PA-MPJPE on 3DPW, 48.8 mm on EMDB, and 41.7 mm on RICH.

更多

依据摘要或发行说明整理,未核验完整正文或独立复现。

追问 / 纠错 / 阅读偏好 · 交给 Agent

评论 · —

昵称自填。评论会展示给其他读者,请勿填写私人信息。

正在读取评论…

OopsieRecover uses abduction to infer the intent behind corrective demonstrations and encodes the knowledge as dense spatial representations and conditional costs. Experiments across six tasks show that only a handful of corrective demonstrations enable a VLA to overcome failures across novel situations without requiring new data for each variation.

更多

依据摘要或发行说明整理,未核验完整正文或独立复现。

追问 / 纠错 / 阅读偏好 · 交给 Agent

评论 · —

昵称自填。评论会展示给其他读者,请勿填写私人信息。

正在读取评论…

VFlash exploits the known structure of visual primitives to predict coordinates jointly using a lightweight drafter that generates an entire primitive in one forward pass. Preliminary measurements indicate approximately 25x faster generation for 3D point motion forecasting relative to autoregressive decoding. A world action model conditioned on VFlash plans achieves 70.2% average success across four MolmoSpaces task suites, compared with 62.6% for the baseline.

更多

依据摘要或发行说明整理,未核验完整正文或独立复现。

追问 / 纠错 / 阅读偏好 · 交给 Agent

评论 · —

昵称自填。评论会展示给其他读者,请勿填写私人信息。

正在读取评论…