Reproduce “Embodied Active Learning of Relational State Abstractions for Bilevel 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
Amber Li, Tom Silver
A framework from MIT CSAIL enables robots to autonomously acquire relational state abstractions, or predicates, for bilevel planning through embodied active learning. This approach strategically interacts with the environment and queries human experts, significantly reducing the number of physical actions and expert interventions needed to solve complex robotic tasks.
Key insights
- Embodied, directed exploration, guided by current model uncertainty, is crucial for efficiently discovering informative states to learn precise predicate boundaries.
- Leveraging an ensemble of neural networks provides a robust measure of epistemic uncertainty, which is vital for guiding both strategic exploration and efficient expert queries.
- The 'planning to learn' paradigm, where a robot actively seeks out and manipulates its environment to generate ambiguous states, significantly accelerates the acquisition of grounded symbols.
- Uncertainty-driven query generation drastically reduces the required human supervision, making the learning process highly sample-efficient for practical robotic applications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.