Tsinghua-led LoopWAM tops RoboChallenge real-robot leaderboard, +83% success via closed-loop self-improvement
机器之心 · wechat · 2026-10-20
LoopWAM, a world action model from Tsinghua, Force Origin, and Jihuan Tech, topped the real-robot RoboChallenge Table30V2 leaderboard on both success rate and process score, beating VLA-DM0.5, and set SOTA on sim benchmarks RoboTwin2.0 (94.5% avg success) and RoboCasa365 (49.5%). All evaluated models were trained on domestic Alibaba Cloud PPU-810E chips, no NVIDIA hardware.
Key ideas:
- Closed-loop self-improvement: success, suboptimal, and failed deployment rollouts all feed video-prediction training; quality steering plus weighted action supervision improves policy quality — no extra expert demos or RL needed.
- Built on Wan2.2-TI2V-5B with a dual video/action branch, jointly denoised with bidirectional cross-attention; source-dependent noise sampling strengthens action-conditioned world modeling.
- Over three deploy-train rounds on Table30V2, success rose 23.67% → 43.33% (83% relative) and process score 36.10 → 56.80.
- On shirt-folding, success climbed from 13.2% to 55.56% through two self-improvement rounds; fruit tidying, sorting, and block retrieval hit 100%.
More from Embodied
- KAIST unveils next-gen humanoid: 12km/h walking, moonwalk, eyes-closed balance — CyberRobooo · 2026-10-20
- AnyGoal: benchmarking world model generalization beyond expert rollouts — galilai-group · 2026-10-20
- EgoXperience: Single-Pass Joint Reconstruction of Camera, Scene, Hands and Objects from Egocentric Video — ropedia-ai · 2026-10-20
- KAIST Humanoid v0.9 pushes 35kg objects while walking using onboard visual feedback — CyberRobooo · 2026-10-20
- DiverseUMI-100 open-sourced: 12k tasks from a 1M+ task dataset scaled in 8 weeks — k7agar · 2026-10-20
- 15 Robotics Research Accounts Worth Following on X, From VLA Datasets to Teleop — FinanceYF5 · 2026-10-20