real-world reinforcement learning for robot policiesExpert
Value as Memory: Online Value-Guided Post-training for General Robot Policies
This preprint from Shanghai Jiao Tong University lets a general robot policy keep learning on the job by feeding it its own recent estimates of task progress and by retraining while it works. On three real long-horizon tasks, including laundry folding, the method raised average success by 29.4 points over the same system without the progress memory and by 32.8 points over conventional batch retraining.