Reproduce “Hierarchical visuomotor control of humanoids” 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
Josh Merel, Arun Ahuja, Vu Pham, Saran Tunyasuvunakool, Siqi Liu, Dhruva Tirumala, Nicolas Heess, Greg Wayne
The authors present a hierarchical controller for a simulated 56-degree-of-freedom humanoid that uses egocentric vision and memory to select motion-capture-based motor skills. Low-level policies track short motions using proprioception, while an LSTM-equipped high-level policy chooses control fragments. In tested tasks, discrete fragment selection enabled Forage maze collection that a flat policy failed to learn, demonstrating performance only in simulation.
Key insights
- Hierarchical control, particularly using fine-grained 'control fragments,' is more effective than end-to-end learning for complex visuomotor tasks in high-DoF humanoids.
- The control fragment method offers a scalable way to create and manage large repertoires of low-level motor skills, reducing the need for laborious manual transition design.
- Integrating memory (via LSTM) into the high-level controller allows agents to exhibit memory-dependent behavior, solving tasks that require retaining past information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.