Learning Robot Policies from Sparse Success Signals via STL-Guided Stein Variational Policy Gradient

TL;DR

This paper introduces STL-SVPG, a method for learning robot policies from sparse success signals. It uses smooth STL robustness as a training objective to assign credit to policy actions. STL-SVPG achieves high success rates on various benchmarks.

Why It Matters

STL-SVPG enables robots to learn complex tasks with sparse feedback, which is crucial for real-world applications. It can handle intricate physical interactions and event-triggered responses. This method has the potential to improve robot performance in various domains.

Key Findings

• STL-SVPG outperforms compared methods on five of six benchmarks
• Simulation-trained policies transfer to real-world tasks
• STL-SVPG handles event-triggered responses and physical interactions