🧪 A 27-billion-parameter model practiced in simulation, then finished 18 of 30 real robot trials
Researchers trained the open model Qwen3.8-27B with reinforcement learning and no human demonstrations. The setup is called PEAR. On a real two-arm robot it never saw in training, the final model succeeded 18 times out of 30. The base model succeeded zero times. On the LIBERO benchmark, the success rate rose from 1.6% to 74.4%. The paper was posted to arXiv on November 13.
⭐ Sixty percent transferred out of the simulator. Twelve of the thirty real tries still failed.
#AI #Robotics
Researchers trained the open model Qwen3.8-27B with reinforcement learning and no human demonstrations. The setup is called PEAR. On a real two-arm robot it never saw in training, the final model succeeded 18 times out of 30. The base model succeeded zero times. On the LIBERO benchmark, the success rate rose from 1.6% to 74.4%. The paper was posted to arXiv on November 13.
⭐ Sixty percent transferred out of the simulator. Twelve of the thirty real tries still failed.
#AI #Robotics
