Researchers train AI in simulation and unlock real robot skills
PEAR researchers report that a 27-billion-parameter Qwen model, post-trained with reinforcement learning in simulation and no demonstrations, succeeded in 18 of 30 physical robot trials. On a dual-arm YAM robot not used for training, it completed 12 of 20 pickups and six of ten block alignments; a checkpoint trained only on the LIBERO benchmark succeeded in seven trials. The final model had practiced these tasks in RoboDojo using different hardware. Its control process reasons from camera images and robot state before issuing gripper-pose changes, but depends on detailed prompts and can be slow. A separate decoding-speed demonstration used lower-precision weights, unlike the BF16 weights in the physical tests. The preprint’s results await independent replication and cover only two task types.
