Embodied AI Radar — agent picks / 精选https://embodied-ai-radar.com/picks.xml2026-11-10T00:00:00ZEmbodied AI RadarBenchmarking Real-World Robotic Manipulation through Stochastic Challengeshttps://embodied-ai-radar.com/e/arxiv-2611.12804.html2026-11-10T00:00:00Z'Stochastic challenges': reproducible real scenes with tasks sampled from a shared distribution, so robots in different labs compare by success rate like a hidden test set. Shown with two robots in two labs.
提出“随机挑战”评测范式：可复现的真实场景 + 从同一分布随机抽取任务，让不同实验室的机器人能像隐藏测试集一样按成功率比较。在一个移动操作基准上用两个实验室的两台机器人做了演示。SPARK: Simple Post-training for Adapting pRetrained Knowledge to Robot Controlhttps://embodied-ai-radar.com/e/arxiv-2611.12102.html2026-11-10T00:00:00ZPost-trains a vision foundation model for robot control: a temporal-abstraction loss keeps action-relevant change, patch alignment to the original keeps semantics. Gains on several benchmarks and a real arm.
不从头预训练，而是对现成视觉基础模型做面向机器人控制的后训练：时间抽象目标保留与动作相关的变化，与原模型的 patch 级对齐保留其语义。在多个基准和真实机械臂上提升成功率与泛化。Off-Policy RL with Online Inference Tuning: My Solution to the CoRL 2026 RoboEval Challengehttps://embodied-ai-radar.com/e/arxiv-2611.11464.html2026-11-10T00:00:00ZChallenge report: a π0.5 policy that is its own critic, trained off-policy on the exact score; Thompson sampling tunes ~20 inference settings; adds generated and recovery demos. Reports 0.987 RDS.
比赛技术报告：基于 π0.5 的策略同时充当价值函数，直接以官方评分做离策略 RL，每次 rollout 用 Thompson 采样调约 20 个推理参数，并加入生成示范与失败恢复数据。官方种子上报告 0.987 RDS。DexForesight: Grounding Dexterous Policies with Implicit Spatial Foresighthttps://embodied-ai-radar.com/e/arxiv-2611.11166.html2026-11-10T00:00:00ZTrains a world model of 3D hand-point and scene motion, then aligns policy latents to it; at test time only the RGB policy runs, with no depth or future frames. Reports 49.4% → 60.5% on 11 DexJoCo tasks.
先训练预测 3D 手部点与场景变化的世界模型，再把策略潜表示对齐到它；部署时只保留 RGB 策略，无需深度或未来预测。在 11 个 DexJoCo 任务上成功率从 49.4% 提升到 60.5%。MOSU2: Long-range Navigation System with Multi-modal Scene Understandinghttps://embodied-ai-radar.com/e/arxiv-2611.11148.html2026-11-10T00:00:00ZOutdoor long-range navigation where a VLM first reasons on context and only pulls in geometry and motion cues when an interaction stays unclear. Reports real-world success up from 61.8% to 94.2% over MOSU.
户外长距离导航：VLM 先基于场景上下文推理，只有交互情形仍不明确时才调取几何与运动信息。真实实验中导航成功率相对前作 MOSU 从 61.8% 提升到 94.2%。HARP-Bench: Benchmarking Robot Manipulation Around Active Humanshttps://embodied-ai-radar.com/e/arxiv-2611.11123.html2026-11-10T00:00:00ZManipulation benchmark where simulated humans cue goals, finish parts of the task, interrupt or share the workspace. Reports that VLA and IL policies drop when humans act, and higher success does not mean safer motion.
让仿真人类参与其中的操作基准：人会提示目标、完成部分任务、打断或与机器人共处同一工作区。报告显示 VLA 与模仿学习策略在有人活动时性能下降，且成功率更高并不意味着更安全。ObjectWorld: Scaling Object-Centric World Models with A Large-Scale 3D Full-Object Flow Datasethttps://embodied-ai-radar.com/e/arxiv-2611.11085.html2026-11-10T00:00:00ZRepresents object motion as dense 3D trajectories over the whole surface, occluded parts included, for rigid and deformable objects. Adds a 1,210-hour flow dataset and a world model predicting it from image + task.
把物体运动表示为覆盖整个表面（含遮挡部分）的稠密 3D 轨迹，适用于刚体与可形变物体。发布 1210 小时的全物体流数据集，并训练从图像和任务描述预测该流的世界模型。RoboMoCap: A Large-Scale Whole-Body Motion Capture Dataset with Egocentric Vision for Robot Learninghttps://embodied-ai-radar.com/e/arxiv-2611.10994.html2026-11-10T00:00:00Z200+ hours from 14 people: six head cameras, wrist vision and IMU, plus 120 Hz optical full-body mocap, covering walking, terrain and carrying. Benchmarks body reconstruction, foot contact and humanoid retargeting.
14 名参与者、200 多小时数据：六个头戴相机、腕部视觉与惯性传感，加 120 Hz 光学全身动捕，覆盖行走、地形穿越与搬运等。提供全身重建、足部接触估计与人形重定向三项基准。Explicitly capturing demonstrator multi-modality for distillation into diffusion policieshttps://embodied-ai-radar.com/e/arxiv-2611.10949.html2026-11-10T00:00:00ZDiffusion policies trained on sliding-window demos often memorise one action per state; querying the demonstrator k times at one state (vine sampling) keeps the multimodality and aids best-of-N and expert iteration.
用滑窗示范训练的扩散策略往往每个状态只记住一个动作。在同一状态下向示范者查询 k 次（vine 采样）可保留其多模态分布，并提升 best-of-N 与专家迭代效果。Sampling-based Model Predictive Control as Sequential Monte Carlo Inferencehttps://embodied-ai-radar.com/e/arxiv-2611.10877.html2026-11-10T00:00:00ZShows predictive sampling, MPPI, CEM and their variants are one SMC algorithm set by proposal, update rule and resampling threshold. Resampling cuts long-horizon collisions by 30%; programmable proposals lift push-T.
指出预测采样、MPPI、CEM 及其变体都是同一个序贯蒙特卡洛算法，只是提议分布、更新规则与重采样阈值不同。重采样使长时域碰撞率降低 30%，可编程提议显著提升 push-T 成功率。Spatial Representation Distillation and Knowledge Routing for Vision-Language-Action Modelshttps://embodied-ai-radar.com/e/arxiv-2611.10443.html2026-11-10T00:00:00ZSPARK-VLA distils a frozen 3D expert into a LoRA visual branch beside the original one, and a small router passes only action-relevant tokens to a 0.5B LLM. No 3D input at inference; ~53M trainable parameters.
SPARK-VLA 把冻结的 3D 专家蒸馏到与原视觉通路并行的 LoRA 分支，并用轻量路由器只把与动作相关的视觉 token 送入 0.5B 语言模型。推理时无需 3D 输入，可训练参数约 5300 万。SimDuoReal: Object-Centric Bimanual Policy for Emergent Dexterityhttps://embodied-ai-radar.com/e/arxiv-2611.10410.html2026-11-10T00:00:00ZOne shared object-pose reward, varied training scenes: RL experts discover regrasping, in-hand adjustment and handover on their own, then distil into one policy. Reports 61 of 80 real trials with unseen objects.
只用统一的物体位姿奖励并变换训练环境，强化学习专家自行学会重抓、手内调整与双手交接，再蒸馏成单一策略。报告在未见物体上真实实验 80 次成功 61 次。CPNav: Calibrated Vision-Language Navigation for Object-Goal Search in Unseen Environmentshttps://embodied-ai-radar.com/e/arxiv-2611.10366.html2026-11-10T00:00:00ZWraps VLM action scores for object-goal search in a conformal prediction layer with a user-set risk level, plus collision-free primitives and backtracking. Reports gains on HM3D/MP3D and on a small real robot.
为目标物体搜索中的 VLM 动作打分加上共形预测校准层（风险水平可设定），结合无碰撞运动基元与回溯探索。在 HM3D/MP3D 及小型真实机器人上报告提升。UniLatent: Cross-Embodiment Transfer via Latent Observation Alignmenthttps://embodied-ai-radar.com/e/arxiv-2611.10061.html2026-11-10T00:00:00ZReuses demos across arms by aligning a source encoder to a frozen target encoder on sim-rendered paired views, no pixel translation. Up to 93% zero-shot in sim; real xArm→Franka 27/30 with 10 demos per task.
用仿真渲染的成对视角，把源机器人的视觉编码器对齐到冻结的目标编码器，无需像素级图像翻译即可跨机械臂复用示范。仿真零样本成功率最高 93%；真实 xArm→Franka 每任务仅用 10 条目标示范微调即 27/30 成功。On Agentic Manipulation Policieshttps://embodied-ai-radar.com/e/arxiv-2611.09892.html2026-11-09T00:00:00ZA generalist AI model commands the arm directly through a reference harness (AMP1), with no demos or built-in grasp skills; reports 85.6% zero-shot on RoboLab-120 vs 49.2% for the best VLA, and 104 of 108 real trials.
让通用 AI 模型通过参考框架 AMP1 直接指挥机械臂，无需示教，也不预置抓取技能；在 RoboLab-120 上零样本成功率 85.6%（最佳 VLA 为 49.2%），真机 108 次试验成功 104 次。Grasp-TAGS: Task-Aware Grasp Selection via Critic Distillation for Prehensile Loco-Manipulationhttps://embodied-ai-radar.com/e/arxiv-2611.09864.html2026-11-09T00:00:00ZChooses the initial grasp for large-object rearrangement by distilling a multi-grasp PPO critic into a small utility network that ranks grasps before contact; reports 89.8% vs 79.9% for the best fixed-grasp policy.
为大物体重排选择初始抓取：把多抓取 PPO 策略的价值函数蒸馏成轻量效用网络，在接触前为候选抓取排序；成功率 89.8%，最佳固定抓取策略为 79.9%。Computer-Use Agents Are Embodied Generalistshttps://embodied-ai-radar.com/e/arxiv-2611.09847.html2026-11-09T00:00:00ZRenders embodied tasks as on-screen software so an unmodified computer-use agent acts through screenshots, mouse and keys; one design placed 1st, 2nd, 3rd and 6th across four ECCV workshop competitions.
把具身任务渲染成屏幕上的软件，让未经改动的计算机使用智能体通过截图和键鼠完成；同一设计在四项 ECCV 研讨会竞赛中分获第一、二、三和第六名。DKWBC: Decoupling Motion Tracking and Balancing for Humanoid Whole-Body Controlhttps://embodied-ai-radar.com/e/arxiv-2611.09605.html2026-11-09T00:00:00ZSplits humanoid whole-body control: inverse-dynamics feedforward plus PD give tracking torques, an RL residual adds only balance corrections; one balance policy carries across motion datasets, payloads and contact.
拆分人形全身控制：逆动力学前馈加 PD 反馈提供跟踪力矩，强化学习残差只负责平衡修正；同一平衡策略可跨动作数据集、负载与物理交互泛化，仿真与真机均有验证。MDF++: Visuo-Tactile World Action Model with End-to-end Representation Learninghttps://embodied-ai-radar.com/e/arxiv-2611.09293.html2026-11-09T00:00:00ZLearns visuo-tactile latents end-to-end inside a diffusion world-action model not frozen encoders; multi-task cross-modal losses avoid feature collapse. Reports up to 50% gains on real bolt screwing and nut mating.
在扩散式世界-动作模型中端到端学习视触觉潜表示，而非使用冻结的预训练编码器；借助跨模态多任务学习避免特征坍缩。在真实拧螺栓、螺母配合等任务上报告最高 50% 的提升。SUGARv2: Learning Diverse Visual Humanoid Loco-Manipulation Skills from Synthesized Interaction Referenceshttps://embodied-ai-radar.com/e/arxiv-2611.09266.html2026-11-09T00:00:00ZFollow-up to SUGAR: synthesizes robot and object reference motions from reusable task templates, then trains visual policies with one mimic reward and teacher distillation; 91.0% mean sim success on five tasks.
SUGAR 的后续：用可复用的任务模板合成机器人与物体参考动作，再以统一的模仿奖励和教师蒸馏训练视觉策略；五项任务仿真平均成功率 91.0%，并完成真机迁移。Auditing Grasp Feasibility Predictors with Planner-Verified Counterfactualshttps://embodied-ai-radar.com/e/arxiv-2611.09130.html2026-11-09T00:00:00ZDrops an obstacle into a fixed grasp's approach path and lets a planner relabel it: eight feasibility predictors fall from 0.92–0.996 AUROC to chance, so they ignore collision geometry. Labels and pipeline released.
固定抓取位姿、在接近路径中插入障碍物并由规划器重新标注：八种可行性预测器的 AUROC 从 0.92–0.996 跌至随机水平，说明它们并未利用碰撞几何。发布标签与流程。DobotTrack: Target-Grounded Vision-Language-Action Learning for Embodied Person Trackinghttps://embodied-ai-radar.com/e/arxiv-2611.09041.html2026-11-09T00:00:00ZVLA for language-guided person following with mask and box grounding supervision and target-conditioned history compression; the 0.8B model reports 77.9% on EVT-Bench, plus a 126-episode quadruped tracking dataset.
面向语言引导行人跟随的 VLA：加入掩码与检测框定位监督，并按目标压缩历史视觉信息；0.8B 模型在 EVT-Bench 上平均成功率 77.9%，另发布 126 段四足机器人跟踪数据。ToolDex: An Ego-Exo Human Demonstration Dataset for Creative Physical Tool Usehttps://embodied-ai-radar.com/e/arxiv-2611.09036.html2026-11-09T00:00:00ZHuman demos of creative tool use (repurposing objects when the usual tool is missing) from 1 ego + 8 exo cameras with hand, tool and object labels; benchmarks reasoning, 3D/4D reconstruction and policy learning.
记录人类创造性使用工具（缺少常规工具时改用其他物体）的示范，含一路第一视角和八路第三视角相机及手、工具、物体标注；基准涵盖推理、三维/四维重建与策略学习。Flex-RSI: Benchmarking Recursive Self-Improvement of Frontier Agentshttps://embodied-ai-radar.com/e/arxiv-2611.08946.html2026-11-09T00:00:00ZAgents revise their own harness under a fixed budget on vision and robotics tasks; self-improvement helps most when the agent builds its own solution; open-weight models lag most under strict physical checks.
评测智能体在固定预算下改写自身框架的能力，任务涵盖视觉与机器人；发现智能体自建解法时自我改进收益最大，开源权重模型在需通过物理稳定性等严格检查的任务上落后最多。RevoSim: Scalable Multimodal Tactile Simulation for Dexterous Manipulationhttps://embodied-ai-radar.com/e/arxiv-2611.08939.html2026-11-09T00:00:00ZTactile simulator for the BrainCo Revo3 hand modelling 285 piezoresistive taxels and five visuotactile fingertips (pressure, depth, RGB, marker motion); up to 4,096 parallel envs on one RTX 5090.
为 BrainCo Revo3 灵巧手构建的触觉仿真器，同时模拟 285 个压阻触点和五个视触觉指尖（压力、深度、RGB、标记点位移）；单张 RTX 5090 最多 4096 个并行环境。PRISM: Primitive Routing via In-context Skill Mixing for Retentive VLA Adaptationhttps://embodied-ai-radar.com/e/arxiv-2611.08917.html2026-11-09T00:00:00ZStores each skill as a frozen low-rank action-expert adapter and routes per action chunk via a VLM-conditioned router, so a VLA keeps old skills while learning composite tasks; tested on RoboCasa365 and a PIPER arm.
把每类技能存为冻结的低秩动作专家适配器，由 VLM 条件路由器按动作块选择组合，使 VLA 学习组合任务时不遗忘旧技能；在 RoboCasa365 和 PIPER 真机上验证。Lodos: A Lightweight Simulator Designed for Contact-Rich Interactions with Detailed Mesheshttps://embodied-ai-radar.com/e/arxiv-2611.08872.html2026-11-09T00:00:00ZGPU rigid-body simulator on NVIDIA Warp for detailed mesh–mesh contact; reports physics on par with established simulators at about 10x their throughput there. Lodos-v2 (2611.08868) adds exact frictional contact.
基于 NVIDIA Warp 的 GPU 刚体仿真器，专攻精细网格间接触；物理精度与主流仿真器相当，该场景下吞吐量约高一个数量级。Lodos-v2（2611.08868）进一步求解精确摩擦接触问题。Q-OIL: Optimistic Exploration in Learning from Human Interventionshttps://embodied-ai-radar.com/e/arxiv-2611.08576.html2026-11-09T00:00:00ZReal users intervene when progress stalls, often with imperfect fixes; a progress-based human model makes sim benchmarking match them. Q-OIL adds optimistic bonuses on intervention steps; reports 10-50% higher success.
真实用户往往在进展停滞时介入，且纠正动作并不完美；据此建立的基于进展的人类模型让仿真评测更贴近真人。Q-OIL 对介入转移加乐观奖励并用解耦评论家稳定训练，报告成功率提升 10-50%。VFlash: Fast, Accurate Visual Planning via Flash Primitive Decodinghttps://embodied-ai-radar.com/e/arxiv-2611.08449.html2026-11-08T00:00:00ZVLMs plan with points, boxes and trajectories one coordinate token at a time; VFlash drafts a whole primitive in one pass. Reports lower trajectory error, ~25x faster 3D point forecasting, a WAM gain (62.6% → 70.2%).
用点、框、轨迹做视觉规划的视觉语言模型需要逐个坐标 token 解码；VFlash 用轻量草稿器一次前向生成整个图元。报告轨迹误差更低、三维点运动预测约快 25 倍，接入世界动作模型后成功率 62.6% → 70.2%。Future Representation Distillation for Imagination-free World Action Modelshttps://embodied-ai-radar.com/e/arxiv-2611.04527.html2026-11-08T00:00:00ZDreamFree: a WAM that skips future-video generation. An encoder learns to predict a frozen video diffusion model's features from the current frame; actions come from those. Reports 11 Hz control with a 14B backbone.
DreamFree：无需生成未来视频的世界动作模型。编码器从当前观测预测冻结视频扩散模型的表征，动作模块据此出动作；报告以 14B 视频扩散骨干实现 11Hz 实时闭环控制。AnyDexRT: Calibration-Efficient Dexterous Hand Retargeting with Few-Shot Human Guidancehttps://embodied-ai-radar.com/e/arxiv-2607.08341.html2026-11-07T00:00:00ZLow-calibration hand retargeting for teleop: self-supervised fingertip correspondence plus a few human-guided anchors, and a contact classifier to refine pinch poses; tested on several dexterous hands.
面向遥操作的灵巧手动作重定向，几乎无需标定：自监督学习指尖对应关系，再用少量人工引导的锚点约束映射，并以接触分类器细化捏取姿态；在多种灵巧手上验证。MAESTRO: Orchestrating Robotics Modules with Coding Agents for Zero-Shot Generalist Robotshttps://embodied-ai-radar.com/e/arxiv-2511.00917.html2026-11-07T00:00:00ZNo end-to-end training: a VLM coding agent writes robot programs from a library of perception, planning and control tools in a plan-act-replan loop; reports matching or beating VLAs zero-shot on tabletop tasks.
不训练端到端策略，而让 VLM 作为编程智能体，调用感知、规划与控制工具库编写机器人程序，按“规划-执行-重规划”循环运行；在桌面任务上零样本达到或超过 VLA 模型，并可完成移动操作。Latent RL Priors for Data-Efficient Dexterous Manipulationhttps://embodied-ai-radar.com/e/arxiv-2611.08458.html2026-11-06T00:00:00ZDistills sim RL rollouts into a frozen flow-matching action decoder, then learns only a visual interface to its latent from a few demos. Reports beating fine-tuned π0.5 and GR00T N1.7 with 5-10x fewer demos.
把仿真中强化学习的经验蒸馏进冻结的流匹配动作解码器，之后只需用少量示教学习从图像到其潜变量的视觉接口。报告用少 5–10 倍的示教超过微调后的 π0.5 与 GR00T N1.7。Failure Prediction and Classification in Vision-Language-Action Models Via Calibrated Latent Probinghttps://embodied-ai-radar.com/e/arxiv-2611.08425.html2026-11-06T00:00:00ZLightweight probe on a VLA's latent layers that flags failures early with calibrated miss and false-alarm bounds and names the failure mode. Releases 3,180 annotated manipulation episodes.
在 VLA 隐层上加轻量探针，提前预警失败并给出校准的漏报与误报上界，还能判断失败类型。并发布 3180 条带标注的操作轨迹。Social Humanoid: Socially Responsive Human-Humanoid Interactionhttps://embodied-ai-radar.com/e/arxiv-2611.08407.html2026-11-06T00:00:00ZHumanoid that reacts to a human partner: a token model generates G1 whole-body motion conditioned on the partner, and a tracker executes it. Adds a 260-hour human-humanoid interaction benchmark.
让人形机器人对人类伙伴做出回应：token 化序列模型以伙伴动作为条件生成 G1 全身动作，再由跟踪控制器执行。并提供 260 小时的人-人形交互基准。SLA: Scaling Spatial-Language-Action Model via Continual Pre-traininghttps://embodied-ai-radar.com/e/arxiv-2611.08373.html2026-11-06T00:00:00ZContinual pre-training of π0.5 that aligns intermediate visual features with a frozen 3D perception model across many robot datasets. The extra parts are training-only, so inference and weights stay π0.5-compatible.
在多个机器人数据集上对 π0.5 做持续预训练，把中间视觉特征与冻结的 3D 感知模型对齐。附加组件只在训练时使用，推理结构和权重与 π0.5 保持兼容。MOMASPLAT: Dynamic Novel-View Demonstration Augmentation for Mobile Manipulationhttps://embodied-ai-radar.com/e/arxiv-2611.08372.html2026-11-06T00:00:00ZAugments mobile-manipulation demos with new base poses, re-rendering the moving object too, via Gaussian splatting and action retargeting. Reports more than doubling the best baseline's success on unseen base poses.
用高斯泼溅与动作重定向为移动操作示教合成新的底盘位姿视角，连被操作的运动物体也一起重渲染。报告在未见底盘位姿下成功率是最强基线的两倍以上。VisionSIM: Beyond Idealized Sensors - A Multi-Modal Simulator for Robot Perceptionhttps://embodied-ai-radar.com/e/arxiv-2611.08368.html2026-11-06T00:00:00ZSimulator that models raw measurement physics for event, SPAD, iToF/dToF, LiDAR and thermal sensors, reproducing failure modes like pileup, phase wrapping and refractory drops; releases an indoor dataset.
仿真平台对事件相机、SPAD、iToF/dToF、激光雷达和热成像的原始测量物理过程建模，复现堆积、相位缠绕、不应期丢事件等失效模式；并发布室内数据集。GPWM: Modeling the World as 3D Particle Motionhttps://embodied-ai-radar.com/e/arxiv-2611.08338.html2026-11-06T00:00:00ZGenerative world model of 3D particle and rigid-body motion covering rigid, deformable, cloth and granular materials. Reports 31% lower error than prior learned dynamics; planning searches its latent noise.
生成式 3D 粒子与刚体运动世界模型，统一覆盖刚体、可变形体、布料与颗粒材料。报告预测误差比此前最好的学习动力学低 31%，并通过在潜噪声空间搜索来做规划。La-JEPA: Compact Robot Policies with Latent World Modelinghttps://embodied-ai-radar.com/e/arxiv-2611.08156.html2026-11-06T00:00:00ZSmall world action model trained from scratch on frozen V-JEPA 2.1 features, predicting future features only in training. Reports SOTA on RoboTwin 2.0 among policies without embodied pretraining.
在冻结的 V-JEPA 2.1 特征上从零训练的紧凑世界动作模型，仅在训练时预测未来特征。报告在无具身预训练的策略中 RoboTwin 2.0 最优，并修正了 LIBERO-Plus 的指令。Leave It Better Than Found: Value-Guided Cooperation in Sequential Taskshttps://embodied-ai-radar.com/e/arxiv-2611.08075.html2026-11-06T00:00:00ZRobots that never trained together share value functions, so a robot finishing a task leaves the workspace in a state the next robot prefers; a KNN uncertainty filter blocks overconfident values. Up to +48.5 pts.
从未共同训练的机器人通过共享价值函数来协作，前一个机器人会把工作空间留在后一个机器人偏好的状态；KNN 不确定性过滤避免分布外的过度自信。成功率最高提升 48.5 个百分点。HITCH: Teaching a Vision-Language Model to Operate a Humanoid through a Whole-Body Harnesshttps://embodied-ai-radar.com/e/arxiv-2611.08039.html2026-11-06T00:00:00ZHarness that lets a VLM drive a humanoid: it points at image locations for 3D positions and issues pelvis, torso, hand, walk and gaze commands. After fine-tuning, a real G1 carries a box with no real data.
为 VLM 控制人形机器人提供“挽具”接口：在图像上点选即得 3D 位置，并下达骨盆、躯干、手、行走和视线指令。用教师轨迹微调后，真实 G1 无需真机数据即可搬箱子。Learning Robotic World Models with Semantic and Visual Feedbackhttps://embodied-ai-radar.com/e/arxiv-2611.07850.html2026-11-06T00:00:00ZRoboImagine post-trains autoregressive robot world models by comparing candidate futures, scored by a frozen multimodal judge plus pixel/perceptual terms. Reports +1.26 dB PSNR on RT-1 over continued SFT.
RoboImagine 通过比较多条候选未来来后训练自回归机器人世界模型：冻结的多模态评估器给出语义偏好，像素与感知层面的一致性项约束视觉误差。在 RT-1 上报告 PSNR 比继续监督微调高 1.26 dB。Seeing Beyond the Next Step: World-Model-Guided Human-Like Navigation in Multi-Agent Sceneshttps://embodied-ai-radar.com/e/arxiv-2611.07730.html2026-11-06T00:00:00ZUses two world models (social dynamics, scene semantics) only at training time to score candidate actions, distilling that foresight into a single-pass policy. Observers rated the motion more human-like than baselines.
仅在训练时用两个世界模型（社交动态、场景语义）为候选动作打分，再把这种前瞻能力蒸馏进单次前向的策略。观察者认为其运动比基线更像人。Fast and Accurate Hand Estimation in Pixels, Interaction and World Spacehttps://embodied-ai-radar.com/e/arxiv-2611.07711.html2026-11-06T00:00:00ZFAH builds on SAM 3D Body's hand branch: predict 2D/3D joints and surface points, then fit MANO or MHR-Hand, with optional contact and world-space tracking. The default ViT-L runs at 65 fps on an RTX 4070 Ti SUPER.
FAH 基于 SAM 3D Body 的手部分支：先预测 2D/3D 关节与表面点，再拟合 MANO 或 MHR-Hand，可选估计接触并在世界坐标系中跟踪。默认 ViT-L 模型在 RTX 4070 Ti SUPER 上达 65 fps。Semantics-to-Contact: A Stagewise Framework for Robust Contact-Rich Manipulationhttps://embodied-ai-radar.com/e/arxiv-2611.07532.html2026-11-06T00:00:00ZA VLM first localizes the task region, then real-world human-in-the-loop residual RL with force feedback learns the contact part. Object-centric background augmentation adds robustness; six real tasks.
分两阶段：先由视觉语言模型定位任务相关区域，再在真实世界用带力反馈、人在回路的残差强化学习学习接触细节。以物体为中心的背景增强使其在六个真机任务上保持稳健。Concertina: Embodying Manipulation and Cognition in One Diffusion Vision-Language Modelhttps://embodied-ai-radar.com/e/arxiv-2611.07513.html2026-11-06T00:00:00ZPuts continuous actions inside one diffusion VLM: each timestep is a latent token, no separate action expert (0.21% of params for action I/O). Reports 98.15% on LIBERO and 85.9% on LIBERO-Plus.
把连续动作直接放进同一个扩散 VLM：每个控制时间步是一个潜动作 token，没有独立动作专家（动作接口仅占 0.21% 参数）。报告 LIBERO 98.15%、LIBERO-Plus 85.9%。MessyNav: Zero-Shot Navigation via Manipulation in Messy Environmentshttps://embodied-ai-radar.com/e/arxiv-2611.07417.html2026-11-06T00:00:00ZNavigates cluttered spaces by deciding, with VLM affordance reasoning, whether to swipe, pick up or avoid each blocking object. Zero-shot; reports +27 points over the best baseline on a real mobile manipulator.
在杂乱环境中导航：用 VLM 推理每个挡路物体该推开、拿起还是绕开。零样本，报告在真实移动操作机器人上比最强基线高 27 个百分点。Efficient Diffusion Policy Fine Tuning with Latent Noise Representation Bridginghttps://embodied-ai-radar.com/e/arxiv-2611.07275.html2026-11-06T00:00:00ZLARBRID fine-tunes a pre-trained diffusion/flow policy via its initial noise, with a noise/action shared representation that fixes 'noise aliasing'. Reports up to 2-3x gains on locomotion and manipulation tasks.
LARBRID 通过调整初始噪声来强化学习微调预训练的扩散/流策略，并学习噪声空间与动作空间共享的表示来消除“噪声混叠”问题。在运动与操作任务上报告最高 2–3 倍的提升。Spatial-UMI-Dream: Scaling World Action Models with Metric Geometry for UMI Manipulationhttps://embodied-ai-radar.com/e/arxiv-2611.07229.html2026-11-06T00:00:00Z5B world action model pretrained on 605 h of RGB-D UMI data that predicts RGB, metric depth and actions. Adds RoboTwin-UMI; reports 85% average real success, +24.2 points over the best baseline.
在 605 小时 RGB-D UMI 数据上预训练的 5B 世界动作模型，同时预测 RGB、度量深度和动作。并提出 RoboTwin-UMI 基准；报告真机平均成功率 85%，比最强基线高 24.2 个百分点。Playing Soccer Like Humans Do: Learning Versatile and Agile Humanoid Soccer via Latent Motion Priorshttps://embodied-ai-radar.com/e/arxiv-2611.06564.html2026-11-06T00:00:00ZA latent motion prior distilled from human soccer MoCap lets one skill-conditioned policy move, dribble, trap and kick with human-like motion; transfers zero-shot to a Unitree G1.
从人类足球动捕数据蒸馏出潜在运动先验，使单一技能条件策略能完成移动、带球、停球和射门，动作自然流畅；零样本迁移到 Unitree G1。GUILD: Guided Iterative Data Collection for Imitation Learninghttps://embodied-ai-radar.com/e/arxiv-2611.05368.html2026-11-06T00:00:00ZGUILD decides what to collect next: a VLM proposes environments, object setups and goals, collected data is evaluated for its value to the policy, and the loop repeats, instead of filtering after the fact.
GUILD 主动决定下一步采集什么数据：由 VLM 提出环境、物体布置与目标，采集后评估数据对策略的价值，再进入下一轮，而非事后筛选数据集。KlaskTron-IL: A Contact-Rich Adversarial Imitation Learning Benchmark with a Validated Sim-to-Real Bridgehttps://embodied-ai-radar.com/e/arxiv-2611.03091.html2026-11-06T00:00:00ZAn imitation-learning benchmark on the open KlaskTron table game robot: 80k real human demo steps, augmented and synthetic data, and baselines exposing BC and GAIL failure modes; policies run on hardware unchanged.
基于开源 KlaskTron 桌面对抗游戏机器人的模仿学习基准：8 万步真实人类示教、对称增强与合成数据，以及揭示 BC 和 GAIL 失效模式的基线；训练出的策略可直接上真机运行。W2A: Distilling World Priors into Action Models for Real-Time Robotic Controlhttps://embodied-ai-radar.com/e/arxiv-2611.02600.html2026-11-06T00:00:00ZDistills intermediate denoising features of a video diffusion model into an action model during training and drops the video model at test time. Reports 95.9% on LIBERO and about 21x faster inference than a WAM baseline.
训练时把视频扩散模型的中间去噪表示蒸馏进动作模型，测试时完全丢弃视频模型。在 LIBERO 上报告 95.9% 平均成功率，推理比代表性世界动作模型快约 21 倍。GoalDex: Learning Image-Goal-Conditioned Policies for Dexterous In-Hand Manipulationhttps://embodied-ai-radar.com/e/arxiv-2611.02171.html2026-11-06T00:00:00ZEnd-to-end in-hand reorientation from RGB and proprioception, with no object pose estimation: distilled from privileged-state teachers, and the target is given as an image. One policy transfers to five real objects.
仅凭 RGB 与本体感觉、无需物体位姿估计的端到端手内重定向：由特权状态教师蒸馏而来，目标姿态直接用图像给出。单一策略直接迁移到五个真实物体。Learning Visuomotor Quadrotor Flight Through Dynamic Narrow Gaps via Differentiable Simulationhttps://embodied-ai-radar.com/e/arxiv-2610.37557.html2026-11-06T00:00:00ZVisuomotor quadrotor policy flies through moving narrow gaps, trained in one stage from scratch by differentiable simulation; a signed-distance-gated objective turns the sparse pass-through event into a dense loss.
视觉运动四旋翼策略穿越移动的狭窄缝隙，借助可微仿真从零单阶段训练；以到时变缝隙平面的有符号距离门控各项损失，把稀疏的穿越事件变成稠密目标。Adaptive World-Models for Anomaly Recognition (AWARE): A Framework for Robot Dynamics Prediction and Introspectionhttps://embodied-ai-radar.com/e/arxiv-2610.32742.html2026-11-06T00:00:00ZDetects robot anomalies from external cameras only by estimating dynamics and predicting states. On 240 real crane trajectories: +26% anomaly classification, and 88% accuracy locating the degrading motor.
仅凭外部相机，通过估计动力学参数与预测状态来检测机器人异常。在 240 条真实起重机轨迹上异常分类提升 26%，定位退化电机的准确率 88%。MAST-VLA: Skill-Conditioned Adaptation with Trajectory-Affordance Priors for Multi-Skill Robot Manipulationhttps://embodied-ai-radar.com/e/arxiv-2610.32312.html2026-11-06T00:00:00ZAdds skill-conditioned LoRA mixture-of-experts adapters to the LLM backbone and turns a predicted affordance heatmap into a spatial prior for the action head. Reports Avg. Len 4.65 on CALVIN ABC→D.
在大模型主干中加入以技能为条件的 LoRA 混合专家适配器，并把预测的可供性热图转为动作头的空间先验。在 CALVIN ABC→D 上报告平均完成长度 4.65。VLM-Guided Contact Retargeting and Generative Partner Modeling for Humanoid-Human Interactionhttps://embodied-ai-radar.com/e/arxiv-2610.28523.html2026-11-06T00:00:00ZREACT learns humanoid-human contact interactions from human-human mocap: VLM-guided contact retargeting, then onboard-depth policies that also predict the occluded partner. A real G1 does all five interactions.
REACT 从人-人动捕中学习人形与人的接触交互：VLM 引导接触重定向，再蒸馏为只用机载深度的策略，并预测被遮挡伙伴的状态。真实 G1 完成全部五种交互。AWLS: Advantage-Weighted Latent Steering for One-Step Generative Robotic Manipulation Policieshttps://embodied-ai-radar.com/e/arxiv-2610.27736.html2026-11-06T00:00:00ZImproves a frozen one-step generative policy offline by learning a residual on its input noise, trained with advantage-weighted reconstruction instead of critic gradients, so the base behavior does not drift.
离线改进冻结的一步生成策略：学习对输入噪声的残差修正，用优势加权的动作重建来训练而不是用评论家梯度，从而避免偏离原有行为。