Body-Aware Semantic Topological Representation for Humanoid Robot Spatial Understanding

Abstract

A humanoid working in human environments requires spatial understanding to know where objects are and which passages its body can physically traverse. Traditional 3D scene graphs address the environment side, but fail humanoids in two ways – per-frame mapping latency grows as the map accumulates and purely environment-centric maps read duckable passages as solid walls. We present a bodyaware spatial representation addressing both limitations. Perframe work is bounded to a local active window, fusing frames into distance fields and condensing them into a sparse freespace graph. Edges retain clearance-versus-height profiles to derive necessary postures (walk, duck, hop, sidestep) and separate observed obstacles from unobserved space. Across 18 timing runs, our per-frame latency remains within a 1.5x band despite 35x map growth, versus 3.3x-42x growth for Hydra. In simulated trials with a Unitree G1, our posture-labeled memory achieves route success rates of 0.78 (24 procedural buildings) and 0.59 (55 room-to-room home tasks), compared to 0.26 and 0.43 for 2D maps. The persistent profiles also handle posttour queries—for modified body dimensions, lost postures, or language goals—without new observation.