TACT-ful system lets humanoids walk stairs at 1 m/s with 15 kg payload

Thanh Ly and 7 co-authors from multiple institutions present TACT-ful, a training system for humanoid locomotion that fuses a multi-channel terrain cost into both a GPU-parallel divergent component of motion (DCM) foothold planner and a dense per-step affordance reward for an asymmetric actor-critic policy trained end-to-end with PPO from depth images. The terrain cost explicitly captures flatness, steepness, and velocity-aware height feasibility—properties missing from single-height signals—allowing the policy to reason about contact planarity and kinematic reachability. A Bézier swing trajectory with adaptive apex bias extends foothold tracking to joint position-and-orientation, using the arc tangent to guide sole orientation through riser crossings and tread landings.

To support payload tasks, TACT-ful introduces a lower-body compliance training procedure where a virtual wrench is injected at a sampled load attachment point, generating physically consistent forces and moments. Wrench-aware compliance targets replace rigid pose penalties, so the policy learns to yield to load-induced perturbations without requiring force sensors. The entire system trains with standard PPO, no distillation, and no teacher-student staging. In simulation, the policy achieves 1.0 m/s on stairs with risers up to 0.20 m and improves payload robustness up to ~15 kg for centered loads and for moment-dominated wrist loads without fine-tuning. A qualitative hardware demonstration on structured terrain is also provided.