Tag
RL can assist robot data collection during demonstrations through a technique called Data Inoculation, improving data quality for both behavior cloning and RL, as shown in a new paper on stacking cubes.
TAD introduces Trajectory-only Adversarial Distillation, which improves vision-language-action policies using only a fixed bank of teacher trajectories—no teacher action scores or task rewards—via a discriminator providing on-policy rewards, boosting OpenVLA-OFT success rates substantially while trailing white-box on-policy distillation.
RynnWorld-Latent is a cross-embodiment video world model that uses 3D-grounded latent actions as a shared conditioning interface, enabling joint pretraining on human videos, robot demonstrations, and simulation rollouts to generate synthetic data that improves robot policy performance on LIBERO and real-world platforms.
论文提出 Just Action Model (JAM)，将视觉-语言-动作模型与世界-动作模型解耦为骨干初始化、未来观测目标与推理机制三个可分离选择，并用仅在训练期使用的 foresight loss 在 LIBERO、RoboTwin 等基准上超越 FastWAM/ImageWAM，同时减少 73% 的训练 FLOPs。
Md Mirajul Islam's PhD dissertation at NC State presents frameworks for Offline Apprenticeship Learning and Imitation Learning from mixed-quality human demonstrations in human-centric settings like classrooms and ICUs, where online trial-and-error is unsafe or infeasible.
UC Berkeley 教授、Physical Intelligence 联合创始人 Sergey Levine 将其春季深度强化学习课程（CS 185/285）全部 25 讲加客座讲座共约 36 小时的录像与作业代码免费公开在 YouTube 和 GitHub，内容涵盖模仿学习、策略梯度、离线 RL、RL for LLMs 等主题，与 ALOHA、π0 等机器人学习方法及当下推理模型训练直接相关。
The paper introduces SCAMPER, a variational system with separate motion and behavior priors for training an embodied virtual rat model that produces natural movements during both goal-directed behavior and random exploration, enabling learning from sparse rewards typical of neuroscience tasks.
DONIC 是首个跨操作类别与人类动作来源联合训练的通用双手灵巧追踪器，结合 PPO 与行为克隆的反馈循环及每手奖励设计，并发布 BiHOI 多视角数据集，在未见物体上无需微调即大幅超越现有 SOTA。
该论文在多样化虚拟世界中系统研究了延续预训练（CPT）的扩展规律，横跨数据量、多样性和模型规模三个轴共120组训练配置，发现数据提升世界内表现、多样性改善相关世界迁移但对远距离世界收益有限，并拟合出经验扩展定律，指出仅靠模仿式延续预训练难以构建通用具身智能体，还需更丰富的训练阶段。
MorphAct 提出一种上下文参数化框架，通过两个耦合的低秩生成器动态调整机器人策略参数，缓解不同交互上下文中相互冲突的参数更新，在 LIBERO、Meta-World 及真实机器人实验中相比多个 VLA 骨干网络最高提升 16.30%。
提出Temporal Flow Policy (TFP)，将物理时间作为机器人策略记忆的缺失维度，使记忆状态随实测时间间隔演化，从而在扩散策略与π₀.₅上实现更准确的时序控制与记忆读出，在LIBERO及真实机器人任务中优于循环与状态空间记忆基线。
WAVE-0 提出一个轻量级潜在世界模型，将 VLA 策略与世界动作模型（WAM）结合，使预测的动作后果直接引导动作生成，在 RoboTwin2.0 和 LIBERO 系列基准上取得 91.47%/98.7% 的成功率，并将 π0.5 的真实任务成功率从 38.3% 提升至 73.3%。
OGPO 框架通过语义关键状态重标注和目标条件过程奖励，将视觉-语言-动作模型的后训练从轨迹级模仿转向关键状态可达性学习，从而提升机器人操作策略在偏离演示轨迹时的恢复能力。实验在 LIBERO 和 MetaWorld 上超越标准监督微调，并在 LIBERO-Plus 及新提出的 LIBERO-ReAct 基准中展现出更强的分布偏移鲁棒性。
提出了 Keyframe Mnemonics，一种自监督关键帧发现方法，用于解决非马尔可夫环境下的行为 cloning 长时程依赖问题，在合成记忆域达到 100% 成功率，并在机器人操作基准上比最强基线平均绝对提升 13.9%。
该论文提出WATCH，一种跨视频预训练方法，通过任务配对的人类演示学习可迁移的世界-动作模型，无需人-机器人配对数据；在机器人后训练后将MimicDroid的上下文内成功率从40.9%提升至68.3%（Cosmos Policy）以及从59.8%提升至77.7%（Cosmos 3）。
A new paper titled "Native Action-Prior Learning from Videos for World Action Models" is shared on Hugging Face Papers, exploring how to learn action priors directly from videos for world action models.
DreamTrue 是一个多视角、跨具身的机器人世界模型，通过离线几何校准与反事实后训练提升动作跟随与物理合理性，在 AgiBot 数据集上将交互缺陷率从 48.12% 降至 6.25%，并获得 AgiBot World Challenge 2026 世界模型赛道第一名。
A new paper, 'Split the Differences, Pool the Rest: Provably Efficient Multi-Objective Imitation Learning', introduces MA-BC, which pools expert demonstrations only where observed actions do not disagree, and provides upper and lower bounds on sample complexity for learning from experts with differing objectives.
Pantheon describes how it built a highly diverse Universal Manipulation Interface (UMI) robotics dataset in eight weeks with 90 operators, custom in-house hardware, and VLM-based freeform annotation, producing over 1 million unique tasks at $10/hour and releasing a 100-hour annotated sample.
MimicX introduces policy-in-the-loop supervision refinement for video-driven humanoid motion tracking, enabling robots to execute human motions (tennis, football juggling, dance, kung fu, parkour) captured from ordinary video demonstrations, including in collision scenes.