Open-Weight 27B Model Learns Real Robot Control via Simulation RL
The notable takeaway is that simulation-only RL can produce transferable robot-control skills without demonstrations, challenging the assumption that robot data collection or proprietary models are prerequisites. Teams building robotics agents with open-weight models can treat RURL-style post-training as a cheaper alternative to teleoperated data pipelines, and the robustness numbers against π0.5 give buyers a concrete benchmark for task-shift scenarios. Researchers comparing open-weight and frontier approaches will find the 74.4% LIBERO result a useful reference point.