Robotics preprint reports performance gains from new policy adaptation method

Researchers report performance increases of up to 2-3× on robotic locomotion and manipulation control problems using LARBRID, a method for adapting pretrained models that generate robot actions. Their arXiv preprint addresses the computational expense and inefficient use of training samples involved in adapting diffusion and flow policies through reinforcement learning. The method learns a shared representation linking a model's initial noise to its actions, then adjusts the initial noise distribution to steer behavior. The authors also derive a theoretical near-optimal guarantee for their framework. The reported gains concern control tasks in a preprint.