Boston Dynamics puts Atlas on wheels to test its motors
Boston Dynamics has replaced Atlas’s feet with wheels in a research platform for testing motors and whole-body control. The US robotics company is using turns, launches and stops to exercise the existing actuators under continuous load. Engineers adapted simulated motor behaviour and learned movement policies, with findings being carried back to the legged robot.
Artificial Intelligence··Midday
Wheels put the existing motors under continuous load
US robotics company Boston Dynamics has tested its Atlas humanoid with wheels in place of feet. The research platform lets engineers examine the motors and control of the whole body through rolling movements. The company calls it a testbed, with launches, turns and stops chosen to exercise the actuators, the components that drive movement, across different combinations of torque and speed.[1], [2]
The only new hardware was a custom adapter connecting an ordinary scooter rim to Atlas’s existing ankle actuator. Controls engineer Xiaoyu (Andy) Wu says that the software model also had to change: the ankle could rotate continuously rather than carry a foot. This joined a limited physical modification with a different movement model for the same part.[1]
Motor measurements reshape simulated training
Wu’s team measured the motors and adjusted their simulated behaviour against the physical robot. It also aligned low-level control settings across software configurations. An idealized simulator can allow a learned policy to exploit movements that the real actuators cannot perform. Motor characterization and tuning addressed that mismatch between the training environment and the hardware.[1]
Existing reinforcement-learning frameworks supplied the starting point. The engineers changed the rewards used in training, the commands given to the robot and its posture. Reinforcement learning uses rewards to shape the controller’s behaviour. In this configuration, smooth rolling emerged while the whole-body controller maintained stability, with motor-model changes accompanying those adjustments to training and posture.[1]
Human movement supplies rolling and jumping trajectories
Human motion-capture recordings supplied some trajectories, especially movements with sliding feet. For sideways dashes, Atlas used wheel steering and rolling in place of a person’s slide. A spinning jump followed a recorded human trajectory. Wu says that movement worked on the first attempt and describes the lower body rolling into position before takeoff; that account supplies no repeated-trial success rate.[1]
The company is adapting lessons from the wheeled platform to the standard legged Atlas configuration. Wu’s next investigation concerns how a learned control policy responds when an actuator fails. The wheel adapter was built by Josh Geating, while colleagues contributed the learning software and assembled the hardware used in the experiment.[1]
Related columns
For more information on this topic, you can read the related columns.
