In the video, Atlas switches between fast rolling on flat ground and legged walking over obstacles and uneven surfaces, and also performs pirouettes and aggressive maneuvers. Andy Wu, a controls engineer for Atlas, trained all the behaviors in simulation, and they ran on the real robot with surprising accuracy, he said.
According to Wu, the conversion was simple: a single adapter that mounts an off-the-shelf scooter rim onto Atlas's existing ankle actuator is the only new part. In software, each wheel is modeled as an "ankle" that can spin continuously, and the reinforcement learning policy responds in real time to sensor feedback and velocity commands to stay stable against disturbances.
The goal, according to the company, is to push the actuators to their limits and test adaptability to unpredictable environments that are hard to model or simulate. Boston Dynamics says a wheeled Atlas is not a product but a research testbed.