New AI 'HumoSlope' lets humanoids walk up 62.7% slopes blind

Humanoid robots struggle on steep slopes because gravity pulls them into slow, crouched gaits. A team of Chinese researchers from Shanghai Jiao Tong University and industry partners now introduces HumoSlope, a two-stage framework that solves this with physics-guided reinforcement learning. Stage I applies a slope-adaptive Zero Moment Point (ZMP) regularizer evaluated on the local inclined support plane rather than the usual world-horizontal reference, giving the robot a terrain-consistent balance prior. Stage I alone still leads to crouching, so Stage II introduces the Biomechanical Slope Gait Adapter (BSGA), which uses extracted macroscopic terrain descriptors as privileged training signals. BSGA dynamically gates soft reward priors to modulate center-of-mass height and lower-limb coordination: uphill calls for hip-dominant propulsion, downhill for knee-oriented braking.

The key innovation: the deployed actor is entirely proprioceptive—no cameras, LiDAR, or exteroceptive sensors needed online. In Sim-to-Real experiments on outdoor grass slopes, HumoSlope achieved blind continuous traversal of slopes up to 62.7% (32.1°), far exceeding previous model-free methods. This demonstrates that physics-guided priors can replace expensive perception pipelines for extreme terrain locomotion. The work, submitted to arXiv in July 2026, opens the door to humanoid robots operating in search-and-rescue, construction, and planetary exploration where slope angles are extreme and sensing is unreliable.