Spot robot uprights heavy tires in 3.7 seconds by combining RL and physics SIM (rai-inst.com)

Researchers at RAI Institute demonstrated a Boston Dynamics Spot variant performing dynamic, whole‑body manipulation to upright a 15 kg tire—completing the maneuver in an average of 5.9 seconds with a best time of 3.7 seconds—by tightly coupling reinforcement learning (RL) locomotion with a physics‑based, sampling controller. Instead of slow, quasistatic grasps, the system exploits timed impacts, leverage, and coordinated arm‑leg-body motions that emerge from physics simulation, allowing the robot to handle loads beyond its nominal gripper lift capacity. Technically, the stack is hierarchical: a low‑level PPO‑trained locomotion policy (IsaacLab) outputs motor torques to maintain balance and execute commands, while task‑level controllers work in a reduced action space (base velocities, orientation, leg/arm commands). For uprighting, a sampling‑based controller runs 32 MuJoCo threads in parallel, sampling spline‑parameterized action sequences to discover multi‑contact strategies; for rolling, a high‑level RL policy (asymmetric actor‑critic, ~24 hours on one GPU) handles complex frictional dynamics. Domain randomization (mass, friction, shape) helps sim‑to‑real transfer. Current limits include motion‑capture state estimation, offboard compute, and modest object generalization—but the work shows that combining RL and fast physics simulation enables fast, forceful, emergent manipulation strategies that expand robot capabilities beyond traditional pick‑and‑place.