Researchers behind the PEAR system trained a 27-billion-parameter model exclusively in simulation before testing it on a physical robot. The model succeeded in 18 of 30 trials, including 12 pickups and six block alignments, while the base model recorded no successes. A checkpoint trained only on the LIBERO simulation benchmark completed seven of 30 trials. The final model analyzed camera images and robot-state data to determine gripper movements. These results, reported in a preprint, have not yet been independently replicated and indicate that simulated practice can transfer some skills to different robotic hardware.
Source: Read the original article


