Reproduce “SLAP: Shortcut Learning for Abstract Planning” 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
Y. Isabel Liu, Bowen Li, Benjamin Eysenbach, Tom Silver
Researchers developed SLAP, a method integrating reinforcement learning with abstract planning to autonomously discover efficient low-level "shortcuts" for robotic tasks. The approach reduced plan lengths by 32% to 73% across various complex environments and achieved a 100% success rate, significantly outperforming pure reinforcement learning methods.
Key insights
- Integrating classical abstract planning with model-free reinforcement learning allows for the discovery of dynamic, efficient low-level behaviors (shortcuts) beyond human-defined skills.
- Learned shortcuts, often involving multi-object interactions, enable substantial reductions in task execution time by replacing sequences of sub-optimal predefined actions.
- Leveraging object-oriented representations from abstract planning allows learned policies to generalize to novel object configurations and counts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.