Reproduce “Sim-to-Real Learning for Humanoid Box Loco-Manipulation” 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
Jeremy Dao, Helei Duan, Alan Fern
Researchers at Oregon State University, with support from NSF and DARPA, developed a sim-to-real reinforcement learning approach to enable the Digit humanoid robot to autonomously perform complex loco-manipulation. The learned controller successfully walks to, picks up, carries, and sets down boxes of various masses and sizes on real hardware, marking the first such demonstration of a fully learned system for this type of task.
Key insights
- Structured guidance, such as pre-defined hand trajectories in the reward function, significantly accelerates learning for complex manipulation skills.
- Employing a "relative" action space (output added to current joint positions) is crucial for efficient learning of manipulation tasks involving large deviations from neutral postures.
- Extensive domain and scenario randomization effectively enables sim-to-real transfer, leading to policies that generalize to varied box properties beyond their training distribution.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.