Guangxiang's Phi-WM1.0 ActEffect: world model supervises robot training then exits deployment, hitting 98.8% on LIBERO
量子位 · wechat · 2026-09-05
Guangxiang Technology, with Tsinghua's Li Shengbo group, released Phi-WM1.0 ActEffect, a physics-native world model that stays only in training: the robot proposes three complete action candidates (feedforward, MIP coarse, refined), the world model predicts outcomes for each and ranks them against real observations, baking "consequence feedback" into policy weights. At deployment the world model exits entirely, keeping inference lightweight.
Key techniques:
- Language stays on the VLA side; the world model only sees vision and actions, learning physical state transitions in a frozen DINOv3 feature space instead of generating realistic frames.
- Gradient truncation on the ranking loss prevents bad candidates from degrading to flatter good ones.
Results: 98.8% on LIBERO, 80.3% on LIBERO-PLUS with seven distribution shifts (vs 51.5% for Fast-WAM), and 67.5% on the 29-DoF RoboCasa-GR1 (9.2 points ahead of ABot-M0). Ablations confirm the feedback's contribution.
Deployment rationale: shifting "thinking" to training avoids inference cost scaling across factory floors. The company's Phi-BotX1 already ran 21.5 hours fault-free on NIO welding jobs, with commercial partnerships underway with leading automakers.
More from Embodied
- Why fleet buyers beat individuals: service consistency, financing, scale — surmenok · 2026-09-05
- Early Cybercab buyers more likely fleet operators, not individual consumers — surmenok · 2026-09-05
- Robotics team signs purpose-built real-world agentic robot — yunta_tsai · 2026-09-05
- RoboTok: Rice and NVIDIA's internet-scale engine retrieves human demo videos for robot learning — CSProfKGD · 2026-09-05
- Tesla's RIM process kills the paint shop, shrinking Cybercab factory footprint ~50% — elonmusk · 2026-09-05
- XDOF, 3 months out of stealth, in talks for Series B at $1.2B valuation — ZeYanjie · 2026-09-05