Caltech researchers teach humanoid robot to autonomously navigate rough terrain over 70m

Researchers from Caltech have unveiled a reinforcement learning framework that lets humanoid robots navigate complex outdoor terrain autonomously over long distances — a significant step toward practical legged locomotion in the real world. The method, described in a preprint submitted to Humanoids 2026, trains perceptive locomotion policies that use reference trajectories dynamically adjusted to match the actual terrain geometry. By synthesizing SE(2)-controllable references inside the RL training loop, the system projects footsteps onto valid footholds and adjusts swing-foot and center-of-mass trajectories accordingly. The resulting policy exposes a clean velocity interface compatible with standard navigation planners, making integration straightforward.

On hardware, the team deployed their policy on the Unitree G1 humanoid robot, paired with a model predictive controller (MPC) and control barrier functions for navigation planning. Without relying on external compute or off-board sensing, the robot achieved over 70 meters of autonomous closed-loop navigation in real-world outdoor environments that included rough, uneven ground and consecutive flights of stairs. This demonstration validates the approach's robustness and hints at near-term applications in disaster response, industrial inspection, and last-mile delivery — anywhere wheeled robots fail but humans (and humanoids) can still tread.