Overview
- Engineers made a minimal hardware change by bolting a single custom adapter and off‑the‑shelf scooter wheels to Atlas’ existing ankle actuators to create a temporary wheeled platform.
- Teams trained whole‑body controllers with reinforcement learning in large‑scale MuJoCo simulations and did not program steering or rolling behaviors by hand.
- To close the sim‑to‑real gap, engineers measured real actuator behavior and tuned simulation motor models and control parameters before transferring policies to the physical robot.
- The wheeled Atlas autonomously produced emergent maneuvers including sideways dashes and a recorded 360‑degree jump that succeeded on its first real‑world attempt.
- Boston Dynamics says the build is a research testbed, not a product, and its findings are being fed back into Atlas’ standard footed configuration to improve actuator models, whole‑body controllers, and reliability for customer deployments.