RoboCoach uses world-model-imagined failures as coaching: 13.3% to 75% success with 150 demos
Tsinghua · hf · 2026-10-01
Tsinghua's RoboCoach treats world models as active coaches for long-horizon robot manipulation. Its Route-Imagine-Diagnose-Improve loop runs reusable skill experts inside a shared action-conditioned world model, using a progress judge to log the first failing subtask; aggregated records decide which demonstrations to collect and which expert adapters to update.
Imagined and deployed success correlate at rho = 0.840 across 22 task-policy pairs on two sim suites and two real robots. With just 150 extra subtask demonstrations, success rises from 13.3% to 75.0% on Franka and 40.0% to 83.8% on AgileX; coached experts transfer to four held-out compositions at 35.0% average success versus 0% for a uniformly-updated shared-policy baseline.
More from Embodied
- Robotics companies aren't taking 'cheapness' seriously enough, argues founder — hudzah · 2026-10-01
- Blogger predicts humanoid robots will soon run hotel services like laundry — CyberRobooo · 2026-10-01
- NVIDIA's Instant NuRec reconstructs a drivable 3DGS world from driving logs in ~1.5 seconds — rsasaki0109 · 2026-10-01
- Yuanyjie Robotics raises seed round to automate restaurant kitchens with embodied AI — 创业邦 · 2026-10-01
- GPT-6 Astra robot evals show strong task decisions but weak physical control — Galbot · 2026-10-01
- GGSD: Game Self-Play Yields Human-Playable Robot Skills That Solve Unseen Tasks Without Retraining — Seungeun Rho · 2026-10-01