HuMBLE preprint demonstrates human-style steering on three humanoid robots

Researchers report a locomotion framework that preserves human gait characteristics while allowing direct steering on three humanoid robots: Boston Dynamics Atlas R1, Atlas D1 and Unitree G1. The HuMBLE preprint describes indoor and outdoor demonstrations and integration into larger robot control systems. The approach first learns motion from human data, then compresses that behavior into a lightweight controller using the robot's own sensed state and a torso-velocity command. Further reinforcement learning expands the commands it can handle while retaining human movement as a style constraint. The authors report benchmarks against controllers trained without human data.