Reproduce “GCNT: Graph-Based Transformer Policies for Morphology-Agnostic Reinforcement Learning” on OpenResearch

About this paper

Yingbo Luo, Meibao Yao, Xueming Xiao

A framework called Graph-Based Transformer Policies (GCNT) enables universal control for robots with diverse morphologies by explicitly decomposing morphological information extraction and inter-limb communication. Researchers from Jilin University and Changchun University of Science and Technology demonstrate that GCNT achieves leading performance on standard benchmarks and robust zero-shot generalization to unseen robot configurations and physical variations.

Key insights

• Explicitly separating morphological information extraction from inter-limb communication via specialized modules (GCN, WL) and a Transformer leads to more effective and robust universal control policies.
• Combining both local (GCN) and global (WL) structural information effectively overcomes the limitations of prior methods that either lost information or failed to leverage morphology adequately.
• Incorporating learnable physical distance embeddings directly into the multi-head self-attention mechanism significantly improves coordination between distant robot limbs, leading to better overall performance.

Summary by alphaXiv.