Reproduce “Abstract Sim2Real through Approximate Information States” 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
Yunfu Deng, Yuhao Li, Josiah P. Hanna
A framework named ASTRA formalizes the abstract Sim2Real problem, which addresses transferring policies from abstract simulators to real robots by explicitly accounting for partial observability induced by state abstraction. ASTRA uses limited real-world data to ground abstract simulator dynamics through history-conditioned corrections, achieving superior performance on tasks like AntMaze (0.90 success) and real NAO robot navigation (73% success).
Key insights
- State abstraction fundamentally induces partial observability in the target domain, causing the Markov property to fail for abstract states and rendering standard Sim2Real methods ineffective.
- Successfully bridging the abstract Sim2Real gap requires history-based methods for simulator grounding and policy learning to effectively account for the induced partial observability.
- Learning a self-predictive latent dynamics model and task-relevant reward prediction, in addition to abstract state correction, creates stable and informative representations crucial for robust policy transfer from abstract simulators.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.