Reproduce “Entity Abstraction in Visual Model-Based Reinforcement Learning” 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
Rishi Veerapaneni, John D. Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua B. Tenenbaum, Sergey Levine
Researchers at UC Berkeley, Stanford, and MIT developed Object-centric Perception, Prediction, and Planning (OP3), a model-based reinforcement learning framework that autonomously discovers and grounds object representations from raw visual observations without supervision. OP3 achieved an 82% accuracy in combinatorial block-stacking tasks, outperforming state-of-the-art video prediction models by a factor of three and competitive with supervised object-centric models, while also enhancing multi-step planning efficiency.
Key insights
- Modeling a physical scene as a composition of distinct entities with local interactions offers substantial advantages for generalizing to novel combinatorial tasks compared to global scene representations.
- Enforcing entity abstraction through symmetric processing allows knowledge about individual object dynamics to transfer effectively across varying numbers and configurations of objects.
- Interactive inference, utilizing temporal continuity and feedback from actions, is critical for robust unsupervised object grounding and disambiguation, especially in complex or ambiguous visual environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.