OpenHLM
“OpenHLM: An Empirical Recipe for Whole-Body Humanoid Loco-Manipulation”
Source→“π0.5 reaches 91% average task progress, PaliGemma drops to 60%, and random initialization collapses to 42%”
Source→“OpenHLM achieved 87.5% task progress, significantly outperforming NVIDIA's GR00T N1.6 (57.5%) and another baseline VLA (48.8%). Crucially, OpenHLM achieved this using 'less than half the total demonstration time'”
Source→“Led the manuscript outline and high-level policy (VLA) design...”
Source→“Listed as corresponding author and core contributor at Tsinghua University and Spirit AI”
Source→“All our loco-manipulation tasks are carried out by a Unitree G1 robot...”
Source→“GR00T N1.6... exhibits weak grasping and fail to track language-specified targets, despite including humanoid data in their pretraining”
Source→“...in line with recent humanoid loco-manipulation stacks [39, 45, 46, 15].”
Source→AI-extracted from podcast / newsletter / paper summaries. May contain errors.