Weekly digest 2026-W45 · Embodied AI Radar

Embodied AI Radar · Weekly 2026-W45 · 具身智能雷达周报

2026-11-02 → 2026-11-08 · 454 new arXiv papers · 71 with code

本周共收录 454 篇 arXiv 新论文，其中 71 篇附代码。

This week in one paragraph. Robot data without robots was the loudest thread: AIRoA-0 reports learning from 16,151 hours of handheld-gripper demos with no robot data, Open-HumanWAM opens a 50,000-hour egocentric corpus and reports a log-linear link between pretraining hours and success, and HALO trains humanoid skills on data synthesised in procedural homes. The second thread was a warning about scores: RoboColosseum's blinded trials at 124 sites find that rankings barely agree across robots, and SimuGuard traces many simulated "policy failures" to physics-engine artifacts. A third, smaller one: open harnesses that turn general models into robot agents (uni-robot-agent, CaPtain, PyRUA). CoRL 2026 opens on Nov 9; the program guide lists all 687 accepted papers by session and poster, 555 with a short takeaway, and lets you save a personal schedule.

一段话看本周。 最突出的主线是"不用机器人也能采到机器人数据"：AIRoA-0 报告仅用 16,151 小时手持夹爪示教、不含任何机器人数据进行学习；Open-HumanWAM 开源 5 万小时第一视角语料，并报告预训练时长与成功率呈对数线性关系；HALO 在程序化生成的家居场景中合成数据来训练人形技能。第二条主线是对评测分数的提醒：RoboColosseum 在 124 个站点做盲测，发现不同机器人上的排名几乎不一致；SimuGuard 则发现许多仿真中的"策略失败"其实源于物理引擎伪影。另有一条较小的线索：把通用模型变成机器人智能体的开源框架（uni-robot-agent、CaPtain、PyRUA）。CoRL 2026 于 11 月 9 日开幕，会议指南按分会场和海报编号列出全部 687 篇录用论文，其中 555 篇附简短要点，还可以保存个人日程。

Editor's picks · 本周精选

Top 8 of 218 new picks, ranked by community signals (Hugging Face upvotes, GitHub stars). 本周新增 218 条精选，按社区信号（Hugging Face 点赞、GitHub 星标）列出前 8 条。

From Pixels to Actions - Towards Native Vision-Language-Action Policies — NEO-VLA drops the visual encoder: images and language evolve in one backbone and a flow-matching action expert reads several depths. Layer probing shows action signals often peak before the final visual layer. / NEO-VLA 去掉独立视觉编码器：图像与语言在同一主干中共同演化，流匹配动作专家从多个深度读取特征。逐层探测显示，对动作关键的信号往往在最终视觉层之前达到峰值。 · details · 详情
RoboColosseum: Evaluating Physical Intelligence at Scale In-The-Wild — Blinded pairwise trials at 124 sites compare VLAs and frontier agents. Rankings barely agree across robots (W = 0.24), and open rankings poorly predict results after fine-tuning; no policy finishes a G1 task. / 在 124 个站点做盲测两两对比，横跨四种机器人评测 VLA 与前沿智能体。不同机器人的排名几乎不一致（Kendall W = 0.24），开箱排名也难以预测微调后的表现；G1 人形上没有策略完成任何任务。 · details · 详情
Open-HumanWAM: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models — Opens a 50,000-hour annotated egocentric video corpus with pipelines and checkpoints. Reports a log-linear law between pretraining hours and robot success, and 84.6% on RoboTwin 2.0 Clean2Random. / 开源 5 万小时带标注的第一视角视频语料及流程和权重。报告预训练时长与机器人成功率之间的对数线性规律，后训练后在 RoboTwin 2.0 Clean2Random 上达到 84.6%。 · details · 详情
Perceptive Humanoid Motion from Simulation — HALO trains humanoid skills on vision-language-motion data synthesised in a thousand procedural homes, so data scales with compute. Reports zero-shot transfer: 91% over 360 trials in six real environments. / HALO 在上千个程序化生成的家庭场景中合成视觉-语言-动作数据来训练人形技能，数据量随算力增长。报告零微调迁移：六个未建图真实环境的 360 次试验成功率 91%。 · details · 详情
RoboJEPA: Scaling Robotic Latent World Models — A JEPA world model trained across 12 robot embodiments; its imagination error follows a power law in compute, so model quality can be predicted beyond the fitted scale. / 跨 12 种机器人形态训练的 JEPA 世界模型；其想象误差随算力服从幂律，因此可以预测超出拟合规模时的模型质量。 · details · 详情
When the Simulator Fails, Not the Policy: Auditing Contact Artifacts in Simulated Robot Benchmarks — Many benchmark 'policy failures' are physics-engine ejections. SimuGuard traced RoboTwin artifacts to one collision constant; changing it cut artifact episodes from 12/100 to none and lifted five tasks 89% → 98%. / 审计仿真基准：不少“策略失败”其实是物理引擎把物体弹飞。SimuGuard 为 LIBERO、RoboTwin、RoboCasa、RoboDojo 提供审计插件，把 RoboTwin 上的问题追溯到碰撞检测的一个常量；仅改动它，异常回合由 100 回合 12 次降到 0，五个容器任务成功率由 89% 升到 98%。 · details · 详情
AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data — Learned from 16,151 hours of handheld gripper demos with no robot data; reports 78% average success over ten tasks versus 27% for the best other VLA, and releases the dataset, weights and post-training recipe. / 从 16,151 小时手持夹爪演示中学习，全程不采机器人数据；报告十项任务平均成功率 78%，其他 VLA 最好为 27%，并开放数据集、权重和后训练方案。 · details · 详情
Pixel: A Co-Designed Full-Stack Robot System for Learning General-Purpose Mobile Manipulation — Hardware, data and learning designed together: a low-cost wheeled humanoid with compliant grippers, teleoperation and robot-free human data collection, one whole-body controller and a shared policy pipeline. / 把硬件、数据和学习放在一起设计：低成本轮式人形、柔顺夹爪，遥操作与免机器人的人类数据采集，统一的全身控制器，以及人类与机器人示范共用的策略学习流程。 · details · 详情

All other picks by topic · 其余精选（按主题）

VLA & Robot Foundation Models · VLA 与机器人基础模型: ALIGN · Asynchronous Robot Policies · Auriga Technical Report · Barrier-Constrained Guidance For Real-time Asynchronous VLA Execution · Can Astra Solve RoboMME without Breaking the Bank? · CaPtain · Concertina · Cross-Controller Transfer via In-Context Behavior Adaptation · Democratizing Robot-Use Agents · Do Agentic Policies Need to Look at Every Step? Geometric Abstraction for Reasoning and Action · Embodied at Sea · Failure Prediction and Classification in Vision-Language-Action Models Via Calibrated Latent Probing · FineCraft · FlowVLA · FR-VLA · From Reasoning to Action · Graph-Reweighted Behavior Cloning with Human Interventions for Precise Robotic Manipulation · HAF-VLA · HeLM · HiFOR · Human-to-Robot Transfer via Bridging Representation and Action Space in VLAs · LAPT-Scale · MAST-VLA · PACER · PLaW-VLA · PyRUA-REAL · REA · REACT · Relational Spatial Alignment for Vision-Language-Action Models · ResMerge · RoboArchitect · RoboDoctor · Say the Same, Act Differently · SLA · STAR · TinyAstral · Trust-Calibrated VLA Policy Post-Training with Evolving Imagination · VFlash · VLA Policies Do Not Know When Not to Act · VLALagBench
Manipulation & Dexterous Hands · 操作与灵巧手: A Tutorial on Shape Servoing of Deformable Objects-Part I · A Tutorial on Shape Servoing of Deformable Objects-Part II · AWLS · BeyondContact · Constraint-Decomposed Impedance Control from Generative Policy for Articulated Object Manipulation · Contact Imitation Interface · Continual Learning for 6-DoF Grasp Synthesis via Experience and Demonstrations · Correct Robots Before They Make Mistakes · CSO · DexFLEX · DexHOT · DexNinja · ESCAPE · FEIJOA · ForceLDM · GoalDex · GRASP · HiFun · Knowing Where to Look · Latent RL Priors for Data-Efficient Dexterous Manipulation · Learning Object-Centric Motion Priors from Human for Robotic Dexterous Manipulation · MobiAgent · MOMASPLAT · Noise-Invariant Spatial Encoding for 3D Manipulation Policies under Depth Corruption · Online Adaptation to Latent Physical Properties for Learning from Demonstration · Phase-Conditioned Affordance Fields for SE(3)-Equivariant Visuomotor Policies · Physics-Aware Object-Centric SE(3) Forecasting for Real-Time Proactive Dynamic Grasping · Pixel-UMI · Reactive Task-Oriented Robot-Human Handovers via Generative Hypothesis Selection · REAP · RoboAuto · Robot-Matched Soft Fingers for UMI Data Collection · RTPolicy · Semantics-to-Contact · Spatial-UMI-Dream · Spectrally Routed Diffusion Policy for Multi-Task Robotic Manipulation · Synthesizing Dexterous Robot Demonstrations from Human Videos via Virtual Try-On · T(R,O) Grasp++ · TAMPAgent
Tactile & Visuotactile Sensing · 触觉与视触觉感知: TF^2-Steering · DeTac · Equivariant Visual-Tactile Diffusion Policy for Contact-Rich Manipulation · Evaluating Tactile Foundation Models for End-to-End Imitation Learning in Contact-Rich Manipulation · How Firm Should a Grasp Be? · PACE · PanoTac · ReflexArc · TacMind
Locomotion & Humanoids · 运动控制与人形机器人: Coordinating Humanoid Loco-Manipulation through Action-Conditioned Latent Foresight · Ctrl-F · DiffWBC · Energy-Based Policy Orchestration · Eye, Robot · FORGE · HIPC · HITCH · HLO-AMP · HoloBoxing · HuG · MAPO · Movens-Zero · Playing Soccer Like Humans Do · Push the Limit · ROOF · SHOES · Social Humanoid · TACOS · VLM-Guided Contact Retargeting and Generative Partner Modeling for Humanoid-Human Interaction · Walk Like This
World Models · 世界模型: Adaptive World-Models for Anomaly Recognition (AWARE) · Box-WAM · Chunk-wise Model-based Policy Optimization · DepthWorld · Does LeWorldModel Preserve Local Action Geometry? · DreamX · Future Representation Distillation for Imagination-free World Action Models · GPWM · Hierarchical World Models with Implicit Dynamics · Interactive World Simulation with Cosmos for Physical AI · Jev with Foresight · La-JEPA · Learning Robotic World Models with Semantic and Visual Feedback · OpenNWM · ORD-WM · pico-WAM · PrimeSim · ReactWorld · Seeing and Imagining as One · Video Prediction Policy 2 · W2A · WorldBridge · Xiaomi Auto World Model 2.0
Navigation & Embodied Agents · 导航与具身智能体: Adaptive Risk-Certified Event-Triggered Replanning for Dynamic Navigation · Aero-VPA · An Agentic Framework for Proactive Robots with Procedural Assistance Memory · Collaborative Planning for UAV-UGV Teams under Operational Constraints via Masked Error-bounded Neuro-Symbolic Diffusion · COOL · CoopUQ · CoSMO · GLST · Learning Visuomotor Quadrotor Flight Through Dynamic Narrow Gaps via Differentiable Simulation · Look Before You Move · MessyNav · OdorSpaces · OpenFlyV2 · PathLift · Seeing Beyond the Next Step · Training Generalist Navigation Policies via Action Self-Distillation · WayTrace
Perception, SLAM & 3D Mapping · 感知、SLAM 与三维建图: AnchorHEC · CoReMatch · Fast and Accurate Hand Estimation in Pixels, Interaction and World Space · FreeLoc · LateMatch · OpenVIO · RoMo-GS
Planning, Control & Safety · 规划、控制与安全: Optimal Control with Learned Critics under Unmodeled State Dependencies · PolicyIK · Safe Streaming Flow Planning by Aligning Sampling Dynamics with Execution Dynamics · VIGOR-MPC
Robot Learning: RL & Imitation · 机器人学习：强化学习与模仿学习: A Balanced Data Diet · AnyRank · Asking Smarter Questions · Behavior Graphs · Beyond Validation Loss · Control-Ready Uncertainty for Trajectory Diffusion · D&A · Efficient Diffusion Policy Fine Tuning with Latent Noise Representation Bridging · ELA · Generative Flow on Symbolic Abstractions for Embodiment Capability Discovery · GSBR · GUILD · Learning Coordinated Visuomotor Box-Pushing from Solo Demonstrations · Learning to Be Different · Leave It Better Than Found · LLMs as MDP Designers · MIRROR · RACE · Recursive Self-Improvement in Physical AI · RoboFQE · Robot-WATCHDOG · Task-Space Imitation Guidance for Efficient Reinforcement Learning · When Robots Outweigh Objects · Where Small Mistakes Matter
Sim-to-Real & Simulators · 仿真与 Sim-to-Real: AuRA · Closing the Dynamics Gap through Self-Directed Robotic Interaction · CoDA · VisionSIM
Datasets & Benchmarks · 数据集与基准: BEAKER · Beyond Infraction Scores · BURGER · ClothDojo · CRONOS · DexArena · HelpMeOut · Is It the Curriculum or Just the Easy Start? Benchmarking Unsupervised Environment Design on Quadrupeds · KlaskTron-IL · MESA · MolmoSpaces-CommonSense · OmniPiano · PyRUA-DataFarm · RMMBench · RoboJev · RoboScope Kit · Rollout then Recycle
Hardware & Open-Source Robots · 硬件与开源机器人: Hardware-as-Regularization · KlaskTron · MABEL · RAIPAL · YOCO

Also added to the index · 本周补充收录

565 earlier papers got a reviewed takeaway this week, 445 of them from the CoRL 2026 program guide. 本周还为 565 篇较早的论文补充了审阅过的要点，其中 445 篇来自 CoRL 2026 会议指南。

Accepted at venues · 已被会议/期刊接收

From the authors' arXiv comments. 根据作者在 arXiv 上的备注。 All papers per conference · 按会议汇总

AAAI 2026 · Learning Object-Centric Motion Priors from Human for Robotic Dexterous Manipulation · project
CoRL 2026 · Mechanistic Interpretability for End-to-End Self-Driving · code · project
CoRL 2026 · HuG: Learning Humanoid Whole-Body Affordance-Aware Grasping with Active Perception · project
CoRL 2026 · Energy-Based Policy Orchestration · project
CoRL 2026 · CoReMatch: Recurrent Continuous Coordinate Refinement for Real-Time Robotic Feature Matching
CoRL 2026 · HLO-AMP: Hierarchical Loco-Manipulation from Object-Aware Motion Priors
CoRL 2026 · MESA: An Evaluation Framework for Compositional, Semantic, and Spatial Generalization in Robotics · code · project
CoRL 2026 · OdorSpaces: Visual-Olfactory Embodied Navigation in 3D Environments
CoRL 2026 · MessyNav: Zero-Shot Navigation via Manipulation in Messy Environments
CoRL 2026 · Transformer-Based Trajectory Prescription for sEMG-Optimizing Robot-Mediated Rehabilitation
CoRL 2026 · Social Humanoid: Socially Responsive Human-Humanoid Interaction · project
ECCV 2026 · iMED: A Multi-Endoscope Dataset for Surgical 3D Perception · code
Humanoids 2026 · MESSPa: A Multi-modal Embodiment Stimulation Shoulder Pad for Preserving the Sense of Agency in a Detachable Robotic Arm
IROS 2026 · Language-Conditioned Behavioral Cloning for Autonomous Endoscope Guidance
IROS 2026 · Mediating Human-Swarm Interactions Through a Robot Helper: A Shepherding-Inspired Approach
NeurIPS 2026 · MotorSense: A Video-EMG Dataset of Motoric Representations for Action Understanding · project
NeurIPS 2026 · PGGT-Loc: Primitive-Grounded Geometry Transformer for Feed-Forward Camera Localization
NeurIPS 2026 · STEER: Route-Aware Adaptive Reasoning for Autonomous Driving
NeurIPS 2026 · Gibbs Variational Last Layers for Long-Horizon Dynamics Learning
NeurIPS 2026 · ForceLDM: A Force-aware Latent Dynamics Model for Contact-Rich Manipulation · code · project
NeurIPS 2026 · Say the Same, Act Differently: Diagnosing and Steering Text-Action Divergence in Reasoning VLAs · project
NeurIPS 2026 · ReSMap: Recasting Satellite Priors for Robust and Accurate Online HD Map Construction · code · project
NeurIPS 2026 · LLMs as MDP Designers: A Dependency-Aware Agentic Framework for Automated Robot Policy Generation
NeurIPS 2026 · BEAKER: An Expert-Curated Benchmark for Embodied Brains in Self-Driving Chemical Laboratories
NeurIPS 2026 · Efficient Diffusion Policy Fine Tuning with Latent Noise Representation Bridging
NeurIPS 2026 · Semantics-to-Contact: A Stagewise Framework for Robust Contact-Rich Manipulation · project
NeurIPS 2026 · Seeing Beyond the Next Step: World-Model-Guided Human-Like Navigation in Multi-Agent Scenes · project
RA-L · A Practical Causal Framework for Autonomous Mobile Robots: From Simulation to Real-World Human Environments
RA-L · Spectrally Routed Diffusion Policy for Multi-Task Robotic Manipulation
RA-L 2022 · GPU-accelerated Incremental Euclidean Distance Transform for Online Motion Planning of Mobile Robots · code
RA-L 2026 · RAIPAL: A Powerful and Agile Humanoid Arm for Human Tasks
WACV 2027 · 3D-BBDA: Backpropagation-Free Black-Box Domain Adaptation for Robust 3D Point Cloud Recognition · project
WACV 2027 · PadLoc: Training-Free Floor-Plan Localization by Sparse Geometric Matching · code
WACV 2027 · Structure from Collision via Hypothesis-Space Reasoning · project

Upvoted on Hugging Face · Hugging Face 热议

Papers from our feed listed on Hugging Face Daily Papers this week, by upvotes. 本周登上 Hugging Face Daily Papers 的收录论文，按点赞数排序。

Rising on GitHub · GitHub 升温榜

Stars gained in the 7 days to 2026-11-08. 截至 2026-11-08 的 7 天内新增星标。

OpenSenseNova/SenseNova-RoboRSI +168 · Has GPT-6-Astra Achieved Embodied Intelligence?
huggingface/lerobot +114 · LeRobot
google-deepmind/mujoco +103 · MuJoCo
HexaFuture/HexaShell +97 · HexaShell: A Universal Harness for Robot Coding Agents in Simulation and the Real World
Ropedia/ropedia-w0 +91 · Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience
Physical-Intelligence/openpi +81 · π0.5: a Vision-Language-Action Model with Open-World Generalization
TheRobotStudio/SO-ARM100 +76 · SO-ARM100 / SO-101
facebookresearch/sam3 +72 · SAM 3: Segment Anything with Concepts

By topic · 分主题统计

VLA & Robot Foundation Models · VLA 与机器人基础模型: 57
Manipulation & Dexterous Hands · 操作与灵巧手: 99
Tactile & Visuotactile Sensing · 触觉与视触觉感知: 23
Locomotion & Humanoids · 运动控制与人形机器人: 43
World Models · 世界模型: 35
Navigation & Embodied Agents · 导航与具身智能体: 34
Perception, SLAM & 3D Mapping · 感知、SLAM 与三维建图: 34
Planning, Control & Safety · 规划、控制与安全: 29
Robot Learning: RL & Imitation · 机器人学习：强化学习与模仿学习: 54
Sim-to-Real & Simulators · 仿真与 Sim-to-Real: 18
Datasets & Benchmarks · 数据集与基准: 34
Hardware & Open-Source Robots · 硬件与开源机器人: 10

By robot type · 分机器人类型统计

Keyword-matched on title and abstract; a paper can count for several types. 按标题与摘要关键词匹配，一篇论文可计入多个类型。

New papers with code · 本周附代码的新论文

VLA & Robot Foundation Models · VLA 与机器人基础模型

Manipulation & Dexterous Hands · 操作与灵巧手

Tactile & Visuotactile Sensing · 触觉与视触觉感知

Locomotion & Humanoids · 运动控制与人形机器人

World Models · 世界模型

Navigation & Embodied Agents · 导航与具身智能体

Perception, SLAM & 3D Mapping · 感知、SLAM 与三维建图

Planning, Control & Safety · 规划、控制与安全

GPU-accelerated Incremental Euclidean Distance Transform for Online Motion Planning of Mobile Robots · code

Robot Learning: RL & Imitation · 机器人学习：强化学习与模仿学习

Sim-to-Real & Simulators · 仿真与 Sim-to-Real

Datasets & Benchmarks · 数据集与基准

---

Auto-tagged by keyword rules; picks are reviewed by the maintainer agent. Website · GitHub · New to the field? Reading path

主题为关键词规则自动标注；精选由维护智能体审阅。刚入门？看入门路线。