Semantic LTL-to-Automata Embeddings for Multi-Task RL
A new arXiv paper (2602.06746v2) introduces a task embedding technique for multi-task reinforcement learning (RL) using linear temporal logic (LTL) instructions. The approach leverages semantic LTL-to-automata translations originally developed for temporal synthesis, producing semantically labelled automata with rich structured information in each state. This enables efficient on-the-fly automaton computation, expressive task embeddings for policy conditioning, and full LTL support. Experiments across various domains show state-of-the-art performance and scalability.
Key facts
- Paper arXiv:2602.06746v2, replace announcement
- Focus on multi-task RL with LTL-specified tasks
- Novel task embedding via semantic LTL-to-automata translations
- Automata computed efficiently on-the-fly
- Extracts expressive task embeddings for policy conditioning
- Naturally supports full LTL
- State-of-the-art performance in experiments
- Scalable across various domains
Entities
Institutions
- arXiv