Embodied AI progress — VLA / dexterous manipulation / sim2real / embodied models / chips / manufacturing / market
Updated every Sat 20:00 HKT
2026-W42 · 10-12 ~ 10-18 →
- {'title': 'Perceive, Predict, and Plan: a Unified World Action Model for Egocentric Dexterous Manipulation', 'summary': 'We introduce P$^{3}$-WAM, a unified world action model for egocentric dexterous manipulation that learns 3D perception, action-conditioned prediction, and goal-directed planning through three compleme', 'url': 'http://arxiv.org/abs/2610.22363v1'}
- {'title': 'ETH Zurich scientists turn WUJI robotic hand into self-propelled mobile manipulator', 'summary': 'Scientists at ETH Zurich have demonstrated how an anthropomorphic robotic hand can use its fingers not only for manipulation but also for locomotion, enabling the hand to crawl, recover from falls and', 'url': 'https://roboticsandautomationnews.com/2026/10/16/eth-zurich-scientists-turn-wuji-robotic-hand-into-self-propelled-mobile-manipulator/105991/'}
- {'title': 'Gradient-based Planning for World Models at Longer Horizons', 'summary': '<!-- twitter --> <meta name="twitter:title" content="Gradient-based Planning for World Models at Longer Horizons" /> <meta name="twitter:card" content="summary_large_image" /> <meta name="twitter:imag', 'url': 'http://bair.berkeley.edu/blog/2026/04/20/grasp/'}
2026-W41 · 10-05 ~ 10-11 →
- VersaCamVLA 与 VioLA 分别从相机配置泛化和人形全身耦合动作切入，推动端到端 VLA 从固定实验设置走向更复杂部署。[端到端VLA]
- DreamTrue 将反事实失败样本与动作忠实的视频预测结合，RoboRSI 则把执行反馈沉淀为可复用修复经验，分别推进世界模型与机器人自我改进。[控制层·世界模型 / 规划层]
- A Balanced Data Diet、USDCraft 与 World Embedding Benchmark 覆盖探索数据、可交互仿真资产和物理表征评测，补齐 sim-to-real 的数据与验证环节。[控制层·RL / 感知层]
2026-W40 · 09-28 ~ 10-04 →
- 世界-动作模型开始从“预测未来”转向“只在训练需要时建模未来”：潜式 WAM 的测试时消融为低延迟 VLA 部署提供了直接依据。[端到端VLA]
- RPG 与 InterEvolve 分别从技能重构和奖励程序演化切入，展示了不更新模型权重也能让机器人在新任务上自我改进的路径。[控制层·BC / 控制层·RL]
- SkeleWAM、World Observer 与 D-JEPA 共同推进世界模型：更紧凑的几何表示、更持久的视野外状态，以及与执行决策对齐的潜空间。[控制层·世界模型]