SP-HLM links sparse poses to whole-body humanoid control
Original title: SP-HLM: Sparse-Pose-Driven Motion Generation and Whole-Body Control for Humanoid Loco-Manipulation
A vision-language-action model predicts five sparse poses, then a tracker combines them with a learned motion prior.
Researchers present SP-HLM, a sparse-pose interface for humanoid loco-manipulation that separates task intent from whole-body coordination. SP-VLA predicts task-relevant sparse poses and finger joints from multimodal observations and language using asynchronous action-chunk inference. SP-MGM reconstructs full-body motion in real time, and SP-GMT tracks motion using those targets plus a latent prior. Simulation and real-robot experiments report better tracking, coordination, and task success than existing approaches.
- The interface uses five task-critical body poses plus a latent motion prior.
- SP-MGM is trained through two-stage distillation for real-time reconstruction.
- A pretrained encoder maps generated motion to a compact latent representation.
- Results and videos are posted on the project page.
From the original article
1 min read at arxiv.orgHumanoid loco-manipulation requires an effective interface between high-level task policies and whole-body control. Existing interfaces typically rely on a single action representation for both task intent and whole-body coordination. This may burden the high-level policy or leave one aspect underspecified.
By Zhaohui Yang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.