The team uses a full-stack system to aid in the robot's ability to learn from humans. "We first train a low-level policy in simulation via reinforcement learning using existing 40-hour human ...