Berkeley's Flip Anything: coding agent self-learns robot flip policies from 24.7% to 95% success
k7agar · x · 2026-11-10
UC Berkeley's CoRL 2026 paper Flip Anything (Shuangyu Xie, Kaiyuan Chen, Ken Goldberg) shows agentic self-learning for robotics: two YAM arms must flip an industrial hand tool face-down using only static RGB-D cameras and joint encoders. A coding agent builds the policy from scratch as ordinary, inspectable code, then iterates — 360 simulated episodes, grouping failures by mechanism, editing source, regression-testing. Over 36 versions it added 20 motion primitives in 5 families (4 flip strategies), raising tune success from 24.7% to 95.0%; nothing is learned inside the policy, all learning lives in the code. Frozen v016 hits 90.8% held-out (88.3% with depth noise/calibration error), and the same simulation-learned code runs on the real rig via a safety-check bridge, flipping a spring clamp in 3m44s and a tape roll in 2m29s — though real runs still fail in ways the simulator rarely shows.
Related event: Berkeley's FlipAnything Lets Robots Teach Themselves to Flip Objects(2 posts)→
More from Embodied
- Indie dev pulls all-nighter assembling pneumatic control box for new MicroFactory bound for England — viktor_vrp · 2026-11-10
- Indie maker all-nighter: assembling MicroFactory control box hours before UK shipment — viktor_vrp · 2026-11-10
- Takashi Ikegami demos android alter3 talking and gesturing at ECOGS 2026 — AnnaCiaunica · 2026-11-10
- Dyna's humanoid Taku handles real laundry tasks smoothly in laundromats and hotels — CyberRobooo · 2026-11-10
- LLM agent directly controls bimanual robot arms: same model plays Minecraft and writes PyTorch — k7agar · 2026-11-10
- Will robotics models run 100x faster in months? A 3-part speedup breakdown, and the latency question — micoolcho · 2026-11-10