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 · 2026-11-13
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。

10 条内容 · 2 篇重点解读 · 8 条简讯

Is Your Robot's Memory Just a Clock? Temporal Shortcuts in Memory-Augmented VLA Policies:Figure 2: Null events break memory, and only through memory. (a) Three null events in one SwingXtimes episode (front camera; first event of three separate rollouts): the robot pauses, rolls its wrist, or grasps the air above the cube and returns, and no object moves. Memory keeps 32 frames sampled uniformly over the episode, each stamped with its time step. In the visible condition the frames of the event en
Figure 2: Null events break memory, and only through memory. (a) Three null events in one SwingXtimes episode (front camera; first event of three separate rollouts): the robot pauses, rolls its wrist, or grasps the air above the cube and returns, and no object moves. Memory keeps 32 frames sampled uniformly over the episode, each stamped with its time step. In the visible condition the frames of t图源:论文作者 / arXiv
arXiv cs.RO精读

Is Your Robot's Memory Just a Clock? Temporal Shortcuts in Memory-Augmented VLA Policies

Chubin Zhang、Bo An

In RoboMME counting tasks, injecting brief pauses or wrist fidgets reduces success rates from 47 to 16 out of 50 episodes on one task and from 44 to 10 on another for released memory VLA policies. The failure is traced to a temporal shortcut where elapsed time predicts progress in demonstrations, and removing time encoding does not cure it because null event frames still occupy slots in time-sampled memory. Indexing memory by events rather than time recovers half or more of the lost success rate without retraining the policy, whereas training with injected pauses fails to remove the failure and impairs counting.

机器人基础模型memory-augmented VLAtemporal shortcuts
论文
展开解读与证据

In RoboMME counting tasks, injecting brief pauses or wrist fidgets reduces success rates from 47 to 16 out of 50 episodes on one task and from 44 to 10 on another for released memory VLA policies. The failure is traced to a temporal shortcut where elapsed time predicts progress in demonstrations, and removing time encoding does not cure it because null event frames still occupy slots in time-sampled memory. Indexing memory by events rather than time recovers half or more of the lost success rate without retraining the policy, whereas training with injected pauses fails to remove the failure and impairs counting.

为什么值得关注

A memory designed to track what happened should advance with events rather than clock ticks to avoid vulnerability to null events in manipulation tasks.

研究问题

Memory-augmented vision-language-action policies that sample frames uniformly by time step fail to distinguish waiting from working, leading to phantom progress where null events like pauses are counted as task completion.

方法原理

The failure is traced to a temporal shortcut where elapsed time predicts progress in demonstrations, and removing time encoding does not cure it because null event frames still occupy slots in time-sampled memory.

结果与证据

In RoboMME counting tasks, injecting brief pauses or wrist fidgets reduces success rates from 47 to 16 out of 50 episodes on one task and from 44 to 10 on another for released memory VLA policies. Indexing memory by events rather than time recovers half or more of the lost success rate without retraining the policy, whereas training with injected pauses fails to remove the failure and impairs counting.

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:Chubin Zhang、Bo An

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…

One Policy for 100 Household Tasks: Our Solution for the 2026 BEHAVIOR Challenge:Figure 1 : The 2026 model. The VLM (left) encodes three camera images, register and scratch tokens, the task and stage embeddings and the robot state; the smaller action expert (right) denoises a 30-step action chunk and reads the VLM through a learnable mixture of the keys and values of all VLM layers. The task and stage embeddings also enter the expert directly through two conditioning tokens. Every attention laye
Figure 1 : The 2026 model. The VLM (left) encodes three camera images, register and scratch tokens, the task and stage embeddings and the robot state; the smaller action expert (right) denoises a 30-step action chunk and reads the VLM through a learnable mixture of the keys and values of all VLM layers. The task and stage embeddings also enter the expert directly through two conditioning tokens. E图源:论文作者 / arXiv
arXiv cs.CV filtered精读

One Policy for 100 Household Tasks: Our Solution for the 2026 BEHAVIOR Challenge

Ilia Larchenko、Gleb Zarin 等 · 7 位作者

The 2026 solution uses a single vision-language-action model checkpoint for all 100 tasks, warm-started from the best 2025 checkpoint and trained for 200k steps on 211 million demonstration frames. On the 50 tasks shared with 2025, the Q score rises from 0.260 to 0.371, and across all 100 tasks the single model scores a Q of 0.392. For 14 difficult tasks routed to a GPT-6 Astra coding agent, the mean Q is 0.639, compared to 0.108 for the learned policy on those same tasks. The public submission achieved a Q of 0.446 with 27.3% full-task success, placing ninth overall and first among independent teams on the final leaderboard.

机器人基础模型BEHAVIOR ChallengeVLA
论文
展开解读与证据

The 2026 solution uses a single vision-language-action model checkpoint for all 100 tasks, warm-started from the best 2025 checkpoint and trained for 200k steps on 211 million demonstration frames. On the 50 tasks shared with 2025, the Q score rises from 0.260 to 0.371, and across all 100 tasks the single model scores a Q of 0.392. For 14 difficult tasks routed to a GPT-6 Astra coding agent, the mean Q is 0.639, compared to 0.108 for the learned policy on those same tasks. The public submission achieved a Q of 0.446 with 27.3% full-task success, placing ninth overall and first among independent teams on the final leaderboard.

为什么值得关注

The results demonstrate that extended training on a unified dataset and hybrid agent-policy routing can significantly improve performance on long-horizon household tasks in simulation.

研究问题

The 2025 solution relied on four separate checkpoints for 50 tasks, and the model was identified as undertrained, limiting its ability to scale to the doubled 100-task set of the 2026 BEHAVIOR Challenge.

方法原理

The 2026 solution uses a single vision-language-action model checkpoint for all 100 tasks, warm-started from the best 2025 checkpoint and trained for 200k steps on 211 million demonstration frames.

结果与证据

On the 50 tasks shared with 2025, the Q score rises from 0.260 to 0.371, and across all 100 tasks the single model scores a Q of 0.392. For 14 difficult tasks routed to a GPT-6 Astra coding agent, the mean Q is 0.639, compared to 0.108 for the learned policy on those same tasks. The public submission achieved a Q of 0.446 with 27.3% full-task success, placing ninth overall and first among independent teams on the final leaderboard.

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:Ilia Larchenko、Gleb Zarin、Atindra Nair、Keerti Manney、Vitaly Kleban、Mirko Kiefer、Anjana Yodaiken

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…

SIGMO is a sigmoid-gated interaction-aware motion planner that uses polynomial sigmoid functions to analytically estimate yield scores of surrounding agents based on physics-grounded features. Validated in interaction-intensive simulations and hardware experiments, SIGMO balances proactive and defensive behavior while remaining computationally lightweight.

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…

TanCRAFT is a four-stage real-time predictive assistance system that tracks scene state, reasons about partial assembly, and forecasts near-future motion to intervene before placement errors occur. In a within-subject user study on tangram assembly, predictive feedback significantly reduced mistakes and yielded the lowest mean perceived workload and shortest completion time compared to unguided sessions.

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…

The study tests a physics-based simulation in NVIDIA Isaac Sim for Healthcare, using Stable Neo-Hookean tissue solved by Vertex Block Descent on one GPU to reproduce vein collapse mechanics. The chosen solver setting reduced time per frame by a factor of 976 to 30.4 ms for the physics solver, with the full stack including rendering taking 56.4 ms per frame.

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…

An affect-adaptive negotiation strategy incorporates CLIFER to personalize valence-arousal perception in real time, using differential feedback to adjust the agent's target utility during live interaction. In a study with 66 participants negotiating with a humanoid robot, the CL-enabled agent achieved significantly higher individual utility and elicited higher concession rates than the FaceChannel 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…

UGG is a unified generative framework for 4D scene understanding that jointly models RGB video and geometric representations like depth and camera pose via conditional completion. The model enables inference-time search over multiple scene hypotheses using geometric evidence from observed RGB frames to improve prediction without ground truth geometry.

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…

Sigma-LeWM introduces backward dynamics and inverse dynamics action predictor losses to LeWorldModel to enforce structure on latent embeddings without a grounding decoder. These additions lead to gains in performance and planning efficiency on standard world model benchmarks and simulated robot manipulation tasks.

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…

BODYinTRANSIT-Lab is a multimodal dataset capturing affective and perceptual responses to the Footsteps Illusion from 101 participants, combining kinematics with physiological signals and subjective ratings. A deep learning benchmark using walking kinematics achieved mean balanced accuracies of 45.9% for affective valence and 44.5% for perceived weight on within-session splits.

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…

The paper extends structural causal bandits to partially known causal structures by developing graphical criteria for minimal intervention sets under Markov equivalence. It derives target reward bounds from source distributions through partial transportability and formulates counterfactual structural causal bandits over actions supported by experimental capabilities.

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…