Reproduce “GSR: Learning Structured Reasoning for Embodied 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
Kewei Hu, Michael Zhang, Wei Ying, Tianhao Liu, Guoqiang Hao, Zimeng Li, Wanchan Yu, Jiajian Jing and 2 more
Shanghai Jiao Tong University's KLM Motion Lab introduced GSR, a framework that employs semantically grounded scene graphs as explicit intermediate representations for embodied manipulation. This approach enables robots to learn structured reasoning, outperforming prompting-based LLM baselines in zero-shot, long-horizon, and complex spatial/causal reasoning tasks on RLBench, LIBERO, and GSR-Bench.
Key insights
- Explicitly modeling world-state evolution through semantically grounded scene graphs provides a strong inductive bias for scalable and robust embodied reasoning.
- Structured intermediate representations effectively disentangle task logic from perceptual variability, enhancing generalization across diverse scenes and tasks.
- Reinforcement fine-tuning using specific reward functions is essential for aligning the LLM's structured reasoning with real-world physical execution constraints and reducing hallucinations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.