RoboCoach uses a world model to target weak subskills, lifting robot success 13.3% to 75%
jiqizhixin · x · 2026-11-08
Renmin University of China, Tsinghua and collaborators propose RoboCoach, which uses a world model to direct human coaching for compositional robot skills, addressing wasted demonstration collection (preprint available).
- Bottleneck: uniform whole-task data collection spends a fixed budget on behaviors the policy already performs reliably, while failures concentrate in a few subskills
- Mechanism: route instructions to reusable skill experts, roll actions forward inside the CoachWorld world model, and use a progress judge to identify the first unfinished subtask; recurring failures become targeted demonstration requests, with humans collecting examples and only the corresponding LoRA expert updated
- Results: across three Franka tasks, success rises from 13.3% to 75.0% after three rounds totaling 150 additional subtask demonstrations
More from Embodied
- Beagle 机器人采收罗马生菜：机器人+AI 种菜成趋势 — RachelVT42 · 2026-11-08
- Tesla to wind down Cybertruck production by mid-2027 as Musk doubles down on Cybercab — XFreeze · 2026-11-08
- Humanoid robots aren't solved: researcher demos social reaction to humans with no mocap, presenting at CoRL — databoydg · 2026-11-08
- NVIDIA's Jim Fan Teases GR00T Demo: Real Hardware, Fully Autonomous, Not Sped Up — McDonaghMatthew · 2026-11-08
- Marc Andreessen says he nearly cracked his phone trying a viral robot spin move — beffjezos · 2026-11-08
- Prediction: GPT-6.2-Level Intelligence at 1,000 Tokens/Sec Next Year, Mass Humanoid Manufacturing — ChrisGPT · 2026-11-08