Why robots should learn without cameras for reliable hands-on tasks

Vision helps robots learn manipulation faster, but makes them brittle—they fail when lighting or angles change. Researchers trained a vision-enabled teacher policy in the real world, then distilled its knowledge into a vision-free student that uses only touch and joint sensors. The student learned the task in 50 minutes and generalized robustly to 8 unseen variants of a NIST assembly benchmark with 95% success, outperforming standard baselines without domain randomization tricks.