Reproduce “Learning Motion Skills with Adaptive Assistive Curriculum Force in Humanoid Robots” 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

Zhanxiang Cao, Yang Zhang, Buqing Nie, Huangxuan Lin, Haoyang Li, Yue Gao

A2CF trains a motion policy alongside a state-conditioned assistive-force policy, then reduces assistance as skills improve. On a simulated 29-degree-of-freedom Unitree G1 EDU, it reaches the maximum rough-terrain walking level at about 4,000 iterations, versus 10,000 for DreamWaQ, while force falls below 0.1 N. Walking and dancing transfer to hardware; backflips remain simulation-only.

Key insights

  • Adaptive, curriculum-based external spatial forces, inspired by human motor learning, can significantly accelerate and stabilize the learning of complex robot motor skills.
  • A dual-agent system that jointly optimizes motion policy and assistive force allows for efficient exploration and guides the robot towards robust, autonomous control.
  • Strategic components such as privileged information for the assistive agent, tailored initial force distributions, and random force masking are vital for improving learning efficiency, preventing over-dependence, and ensuring successful sim-to-real transfer.

Summary by alphaXiv.

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