Visuomotor Pick

Key Insight

Rather than dividing a manipulation task into separate perception, planning, and control steps, a visuomotor policy maps camera images directly to joint motor commands. Training such policies end-to-end using reinforcement learning inside simulators like MuJoCo allows the robot to adapt to visual variations and contact dynamics. This direct connection between sight and action enables highly reactive pick-and-place behaviors that can handle moving targets and unexpected physical disturbances during execution.