Reproduce “Learning from Imperfect Demonstrations via Temporal Behavior Tree-Guided Trajectory Repair” 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
Aniruddh G. Puranic, Sebastian Schirmer, John S. Baras, Calin Belta
A formal framework is presented for training robot control policies by repairing imperfect demonstrations using Temporal Behavior Trees (TBTs) and extracting kinematic-model-agnostic potential functions for reward shaping in reinforcement learning. This method achieved up to 47% lower collision costs in multi-agent tasks and demonstrated faster convergence in various navigation scenarios compared to baseline approaches.
Key insights
- Formal repair of imperfect demonstrations using TBTs can effectively pre-process suboptimal data, making it usable for robust and data-efficient policy learning.
- Potential functions derived from the poses of formally repaired trajectories offer a kinematic-model-agnostic reward shaping mechanism, enhancing generalizability across diverse robot platforms.
- Integrating formal methods for data repair with reinforcement learning significantly improves the safety, reliability, and convergence speed of learned policies by guiding agents towards specification-consistent behaviors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.