Berkeley researcher launches RPG: guided self-improvement for embodied agents
berkeley_ai · x · 2026-10-08
Berkeley AI researcher Haozhe Jiang released RPG (Reconstruct, Practice, Go Real), a self-improving embodied agent project, with a blog on motivations and lessons.
Key points
- Agentic robots went viral after GPT-6 (driving, peeling cucumbers, folding towels), but remain slow, costly, and brittle — folding one towel takes 2 minutes and fails on a 45° rotation.
- He questions why bimanual manipulation remains unsolved: hand-written scripts are reliable but don't generalize; foundation models should make capability acquisition cheaper and faster.
- The post explores which emerging foundation-model capabilities contribute to robotic systems and which offer new scaling dimensions.
Related event: Berkeley's RPG Embodied Agent Aces 30 Real-World Tasks(2 posts)→
More from Embodied
- Researchers Put GPT, Claude and Grok Behind the Wheel of a Real Corolla — Only One Succeeded — nordicinst · 2026-10-08
- Mila's QWM world model controls heterogeneous quadrupeds zero-shot, no fine-tuning — Mila_Quebec · 2026-10-08
- Surface Laptop Ultra ships Oct 16 at $2,599 with magnetic USB-C 'Magnetic Connect' charging — tomwarren · 2026-10-08
- Surface Laptop Ultra launches Oct 16 starting at $2,599 with Nvidia RTX Spark — tomwarren · 2026-10-08
- Atomic Machines builds an AI 'matter compiler' to make index-finger-sized circuit breakers — ZoubinGhahrama1 · 2026-10-08
- Jaguar Type 01 debuts on NVIDIA Hyperion platform, software passed 150,000 tests — nvidia · 2026-10-08