Reproduce “Perception-and-action system for humanoid robot task execution in construction” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yanxi Liu, Yizhi Liu

A Vision-based Perception-and-Action (VPA) system allows humanoid robots to learn construction tasks from human demonstrations by translating human motion into humanoid-compatible poses and executing them with stable whole-body control. This system achieved an average MPJPE of 82.45 mm for various construction-related actions on a Unitree G1 robot.

Key insights

  • A combined triplet, reconstruction, and latent-consistency loss function effectively addresses morphological mismatches for accurate human-to-humanoid pose retargeting.
  • Physics-aware reward function design within a teacher-student reinforcement learning framework enables robust whole-body control for dynamically stable execution of construction-specific actions.
  • Distilling a comprehensive teacher policy into a student policy using only proprioceptive states facilitates reliable sim-to-real transfer for humanoid robots performing complex tasks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.