SP-HLM: Sparse-Pose-Driven Motion Generation and Whole-Body Control for Humanoid Loco-Manipulation
Five task-critical body poses, paired with a learned motion prior, guide a humanoid’s whole-body actions.
A vision-language-action model predicts sparse pose targets and finger positions from observations and language instructions. A motion generator reconstructs coordinated full-body movement in real time, which is encoded into a compact motion prior. The controller combines that prior with the explicit pose targets. The authors report better tracking, coordination, and loco-manipulation success than existing approaches in simulation and on a real humanoid, but the abstract gives no numerical gains.
ArXiv · AI/CL/LG's note
A vision-language-action model predicts sparse pose targets and finger positions from observations and language instructions. A motion generator reconstructs coordinated full-body movement in real time, which is encoded into a compact motion prior. The controller combines that prior with the explicit pose targets. The authors report better tracking, coordination, and loco-manipulation success than existing approaches in simulation and on a real humanoid, but the abstract gives no numerical gains.
ArXiv · AI/CL/LG's note
score 5