Boston Dynamics Swaps Atlas Feet for Wheels, Lands Full-Spin Jump

How the control system was trained

Researchers trained the controller through repeated simulation, tuning digital models of the actuators until their response came close to that of the hardware in the robot. The team also used recordings of human movements, including jumping, spinning and sliding, to guide the training process. The goal was to measure how well software trained in simulation can control the motors when the movements are carried out in the real world.

Why wheels instead of feet

Wheels offer a way to test the actuators under conditions that combine speed, torque and weight bearing during motion. They also expose the gaps between how a control system performs in simulation and how it performs on the ground.

Boston Dynamics stresses that the aim is not to replace the two-legged version of Atlas. Instead, the company wants to use the findings to improve control of the main model and make it better at coping with faults and changes during movement.

The experiment on its own does not show that Atlas can carry out autonomous work in factories. For now, it is a research result on how well simulation-trained controllers transfer to physical hardware.