Reproduce “SafeAPT: Safe Simulation-to-Real Robot Learning using Diverse Policies Learned in Simulation” 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
Rituraj Kaushik, Karol Arndt, Ville Kyrki
The SafeAPT framework enables robots to safely and data-efficiently adapt to new tasks in the real world. It generates a diverse policy repertoire in simulation under varying dynamics and employs online learned reward and safety transformation models to minimize safety violations during real-world interaction.
Key insights
- Generating a policy repertoire under a distribution of diverse simulated dynamics significantly enhances safety and robustness when adapting to unknown real-world dynamics.
- Online learning of real-world specific reward and safety transformation models via Gaussian Processes, initialized with simulated priors, is crucial for data-efficient and safe adaptation.
- A safety-aware Bayesian Optimization acquisition function (ESI), incorporating a Lower Confidence Bound on predicted safety, effectively guides policy selection to minimize safety violations during real-world learning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.