Researchers from the RAI Institute, Boston Dynamics and Carnegie Mellon University report HuMBLE, a humanoid walking controller trained to reproduce human walking patterns while following live commands for speed and direction. In operation, it uses a robot's onboard movement and joint readings plus a steering command to set joint targets, rather than replaying the human recordings. In simulated tests on Boston Dynamics' Atlas R1 and Atlas D1 and Unitree's G1, HuMBLE's errors in following movement commands were similar to those of a controller trained without human motion across most tested speeds and directions, though that controller responded faster at high speeds.
A controller trained only to follow movement commands can produce a stiff walk, while one trained to copy human walks may struggle with directions or speeds missing from its examples. HuMBLE aims to retain patterns from those walks while extending the commands the robot can handle.
The team first trained a larger policy using reinforcement learning, a trial-and-reward method, to track full-body human-motion references, then distilled it into a smaller controller. They fine-tuned that controller with two objectives: follow a wider range of steering commands over longer periods and briefly imitate recorded walks. This approach expands the steering range while retaining movement patterns from the demonstrations.
On Atlas R1, HuMBLE's movement included arm swing, knee extension and heel-to-toe foot movement that the researchers say were less evident in Boston Dynamics' default controller and a separate controller trained without human-motion data. The comparison system responded faster to high-speed commands; HuMBLE's transitions were more gradual. Its largest errors came from fast diagonal commands that were sparse or absent in the training data.
The team tested command tracking in simulation and on physical hardware across all three platforms, and demonstrated joystick control indoors and outdoors. In warehouse navigation, a separate system planned waypoints around obstacles and sent walking commands to HuMBLE; the box-parkour demonstration paired it with separate controllers for climbing up and down. HuMBLE handled the walking part, not the full tasks alone.
These results come from a new arXiv preprint, not a peer-reviewed study. The team adjusted recorded human walks to each robot's body proportions and joints, then compared the robot's movements with those adjusted motions. This motion-fidelity score is a proxy for style, not a rating from people. The authors describe the work as standard locomotion on flat terrain; rougher ground, climbing and natural-language steering remain open challenges.
Sources
- HuMBLE paper, arXiv:2610.10489v1 (posted Oct. 7, 2026)
- Cover image: Figure 1 by Mike Zhang et al.; CC BY 4.0.