Berkeley's RPG lets robots watch demos and practice to improve, hitting 30/30 on real tasks
berkeley_ai · x · 2026-10-08
UC Berkeley researchers introduced RPG (Guided Self-Improvement for Embodied Agents), a real-to-sim-to-real pipeline that lets robots decide what to practice and improve through practice.
Pipeline: Demo → Reconstruct → Guided Practice → Self-Improve → Go Real:
- Reconstructs practice tasks from real-world demonstrations
- Uses demo videos to guide skill improvement
- Continually updates a shared skill system from failures
Results:
- 30/30 successes across 3 tasks on real robots
- 95.0% success across 22 manipulation tasks in simulation after 15 rounds of self-improvement
More from Embodied
- MeckaAI raises $60M Series B led by Sequoia for physical AI data layer — philhchen · 2026-10-08
- Europe's industrial robot installations fell 8% to 85,000 in 2024 — ingliguori · 2026-10-08
- TEMPO adds temporal context to VLAs, lifting robot bottle handover success from 44% to 74% — _krishna_murthy · 2026-10-08
- Porting LIBERO to MuJoCo Warp: 130 robot tasks on one $700 AMD GPU — one_does_not_just · 2026-10-08
- A Humanoid Robot Should Build Houses, Garden, and Repair Itself, Says Viral Thread — repligate · 2026-10-08
- Liquid AI shows 10 live-camera d1-3B demos at one forward pass per frame — JosephJacks_ · 2026-10-08