RoboCoach: Active Error Correction via World Models for Embodied RSI
Overview
Pieter Abbeel and his team introduce SuperMimic, a method that enables fast and cost-effective training of full-body humanoid control policies. By using a faster world model and more efficient backpropagation through time, they achieve results previously requiring 128 GPUs over 7 days with just one GPU in 3 hours.
From Pieter Abbeel — X Robotics
Event evidence and corrections
2 attributed source owners. Ownership does not establish independent confirmation. Quantities are reported separately and are never added together.
Reported quantity · Reported success rate: 13.3 percent · Basis not reported · Differing source assertions
“RoboCoach 在有限人工示范预算下，将 Franka 的成功率从 13.3% 提升至 75.0%，AgileX 从 40.0% 提升至 83.8%，相比起基线方法有显著效率提升。”
Exact source · revision 1jiqizhixin
Reported quantity · Reported success rate: 75 percent · Basis not reported · Differing source assertions
“RoboCoach 在有限人工示范预算下，将 Franka 的成功率从 13.3% 提升至 75.0%，AgileX 从 40.0% 提升至 83.8%，相比起基线方法有显著效率提升。”
Exact source · revision 1jiqizhixin
Reported quantity · GPU-hours saved: 1000 other · Basis not reported
“【引用 @qiayuanliao】We found a way to scale humanoid control that reduces GPU-hours by more than 1000×.”
Exact source · revision 1berkeley
Reported quantity · motions processed: 129514 other · Basis not reported
“Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.”
Exact source · revision 1berkeley
Reported quantity · training time: 3 hours · Basis not reported
“Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.”
Exact source · revision 1berkeley
Reported quantity · GPU used: 1 other · Basis not reported
“Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.”
Exact source · revision 1berkeley
Reported quantity · state of the art GPU-hours: 128 other · Basis not reported
“Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.”
Exact source · revision 1berkeley
Reported quantity · state of the art days: 7 other · Basis not reported
“Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.”
Exact source · revision 1berkeley
Reported quantity · state of the art GPUs: 128 other · Basis not reported
“Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.”
Exact source · revision 1berkeley
Evidence dependency changes (4)
- 2026-10-30T21:41:56.775Z · evidence updated · source revision 1. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T21:41:40.503Z · evidence updated · source revision 1. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T20:04:32.368Z · evidence updated · source revision 1. Evidence extraction was updated; current source attributions are shown above.
- 2026-10-30T20:04:16.611Z · evidence updated · source revision 1. Evidence extraction was updated; current source attributions are shown above.
Developments
- 2026-10-26 04:15 UTC · 1 reportsRoboCoach: Active Error Correction via World Models for Embodied RSIMachine Heart — Robotics on WeChat：RoboCoach: World Models as Active Coaches for Compositional Robot Skills
- 2026-10-21 16:55 UTC · 1 reportsSuperMimic: Fast and Cheap Humanoid Control TrainingPieter Abbeel — X Robotics：Game changer for humanoid full-body control training
Report timeline
Follow attributed reports and material updates.
- Machine Heart — Robotics on WeChatSignalRoboCoach: World Models as Active Coaches for Compositional Robot Skills
RoboCoach is a system that uses world models to enable efficient active correction in embodied reinforcement learning. By decomposing long tasks into subtasks and directing human demonstrations to specific experts, RoboCoach achieves significant performance improvements on real robots like Franka and AgileX, with success rates increasing from 13.3% to 75.0% and 40.0% to 83.8%, respectively. The system's RIDI framework (Route—Imagine—Diagnose—Improve) enables targeted skill refinement, reducing the need for extensive human intervention.
- Pieter Abbeel — X RoboticsSignalGame changer for humanoid full-body control training
Pieter Abbeel and his team introduce SuperMimic, a method that enables fast and cost-effective training of full-body humanoid control policies. By using a faster world model and more efficient backpropagation through time, they achieve results previously requiring 128 GPUs over 7 days with just one GPU in 3 hours.
Timezone · UTC
Article dates follow your selected timezone. Briefing editions use Hong Kong time (UTC+8).