Reproduce “Hierarchical Planning for Long-Horizon Manipulation with Geometric and Symbolic Scene Graphs” 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

Yifeng Zhu, Jonathan Tremblay, Stan Birchfield, Yuke Zhu

Researchers from NVIDIA and The University of Texas at Austin introduce a visually-grounded hierarchical planning framework that combines neuro-symbolic task planning with graph-based motion generation for complex, long-horizon manipulation tasks. This approach achieves a four orders of magnitude speedup in task planning over traditional methods and demonstrates robust performance with up to 70.6% overall success on a real robot operating directly from visual input.

Key insights

  • Integrating neuro-symbolic task planning with graph-based motion generation effectively bridges the gap between discrete symbolic reasoning and continuous robot control, enabling strong generalization and efficiency.
  • Hierarchical scene graphs provide a structured, object-centric abstraction that allows GNNs to leverage relational inductive biases, crucial for generalizing to varying scene compositions and novel tasks.
  • The Regression Planning Network's backward planning mechanism significantly improves performance for long-horizon tasks by recursively predicting previous subgoals, overcoming limitations of direct subgoal prediction.

Summary by alphaXiv.

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