Researchers have developed HumanoidPose, a novel framework designed to estimate the full-body joint configurations of humanoid robots solely from video input. This system treats visual data as a form of proprioception, combining pre-trained human-pose features with robot kinematics to predict joint positions. The framework can retarget human motions to sixteen different humanoid embodiments, enabling learning from synthetic data, and further improves real-world generalization through self-training on unlabeled online videos. HumanoidPose has demonstrated significant improvements in joint-angle and position accuracy compared to existing methods on real humanoid platforms. AI
IMPACT This research could enable robots to better understand their own state and movements from visual input, potentially improving human-robot interaction and robot learning from video data.
RANK_REASON The cluster contains an academic paper detailing a new research framework. [lever_c_demoted from research: ic=1 ai=1.0]
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
