Event

PKU-Lingchu RoboFFT Accepted to CoRL 2026

1 reports · 1 sources

RoboFFT, a generative robot policy reinforcement learning post-training method proposed by the PKU-Lingchu Joint Laboratory, has been accepted to CoRL 2026. The method reuses the forward noising process from diffusion model and flow model pretraining to construct an approximate probability ratio, which is used for PPO-style clipped updates. The research adopts an offline-to-online pipeline: first, a flow policy is pretrained using human teleoperation demonstrations; then, offline reinforcement learning and imitation learning are performed using robot-collected trajectories; finally, updates are made through online interaction; task feedback is judged as success or failure by humans. Validation covers simulation tasks, a vision-language-action model, and four real-robot tasks: peg insertion, eggplant grasping, pouring, and retrieving objects from a cabinet, with each real-robot task tested 20 times.