Fundamental Performance Limits for Sensor-Based Robot Control and Policy Learning