AI wrote its own robot flipping tricks, reaching 95% in simulation. Real objects were harder.

A pair of robot arms can turn over a hammer by passing it between hands, or flip a wide object by pushing it onto its side and toppling it. In Berkeley's FlipAnything study, a coding agent developed and revised the program that selects such moves. The team's November 9 preprint reports an increase from 24.7% to 95% success in a physics simulator across 24 objects.

Turning something over requires more than grasping and rotating. The robot must find a reachable grip, avoid losing contact and recognize when the object has landed the wrong way. Researchers supplied an English task description, object models and the robot's mechanical description. The agent then tested its generated controller, examined failures and revised the executable logic. The accumulated skill resides in inspectable code rather than a newly trained action model. During execution, the controller uses camera observations and the robot's own joint information.

The initial handover-only program succeeded in 89 of 360 simulated trials. Adding strategies such as toppling raised that to 290; further improvements to gripping, contact handling and recovery produced 342 successes in the best tuned controller. Its responses included moving an object before attempting a handover and adding a push when a flip stopped halfway. For inspection and assembly, the interesting possibility is a development process that turns failed practice into a richer robot program.

The 95% figure comes from the repeatedly optimized simulation set, not a factory deployment or an unlimited assortment of objects. Physical trials on three objects absent from simulated learning were much less convincing: two successes in three attempts for a clamp, three of three for blue tape and one of three for a folding screwdriver. The project page's initial commit dates to October 26; the November 9 preprint documents the study. Reliable real-world use still requires better perception, contact handling and compatibility with the robot's physical limits.

Independently rewritten from the CC BY 4.0 paper by Shuangyu Xie, Kaiyuan Chen and Ken Goldberg; no source images reproduced. License: https://creativecommons.org/licenses/by/4.0/.

Original source times: arXiv paper, Nov 9, 2026, 14:09:55 Beijing time / 1:09:55 AM EST; initial GitHub project-page commit, Oct 26, 2026, 15:26:58 Beijing time / 3:26:58 AM EDT.