Online Signal Temporal Logic Tree Search for Guided Imitation Learning in Stochastic Domains
Follow The Rules: Online Signal Temporal Logic Tree Search for Guided Imitation Learning in Stochastic Domains, ICRA 2023
Jasmine Jerry Aloor, Jay Patrikar*, Parv Kapoor, Jean Oh and Sebastian Scherer
Paper: https://arxiv.org/abs/2209.13737
GitHub: https://github.com/castacks/mcts-stl-...
A challenge to deploying robots in the real world is to ensure they seamlessly integrate real-life rules into various robot learning policies. Learning-from-Demonstration (LfD) policies trained on real-world data often fail to distill the underlying rules due to imperfect and incorrect demonstrations. In this work, we explore encoding rules as high-level task specifications to improve the online performance of the LfD policies.
This paper presents one of the first methods that combines high-level rules using Signal Temporal Logic (STL) into the policies learned from demonstrations through Monte-Carlo Tree Search. We evaluate our algorithm on the real-world problem of imitation learning for autonomous general aviation aircraft in a variety of missions. Our method showcases 60% improved performance over baseline LfD methods that do not use STL heuristics.
#robotics #deeplearning #ai #artificailintelligence #safety #uam #aam #LfD #advancedairmobility #vtol #evtol #autonomy #learning #icra2023