Reproduce “Data-efficient Domain Randomization with Bayesian Optimization” 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
Fabio Muratore, Christian Eilers, Michael Gienger, Jan Peters
Bayesian Domain Randomization (BayRn) is an algorithm that adaptively adjusts simulation parameter distributions based on sparse real-world returns, leveraging Bayesian Optimization to close the sim-to-real gap. This method enables robust policy transfer for robotic tasks, outperforming static domain randomization and mitigating issues observed in explicit system identification approaches.
Key insights
- BayRn implicitly identifies robust domain parameter distributions by directly optimizing for real-world policy performance, bypassing the complexities and potential instabilities of explicit system identification.
- Bayesian Optimization's sample efficiency significantly reduces the number of costly real-world interactions required for adaptive sim-to-real transfer.
- Adapting the domain parameter distribution through real-world feedback leads to more robust and consistently performing policies on physical systems compared to fixed, handcrafted randomization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.