Zero-Shot STL Planning with Disjunctive Branch Selection in Dynamic Semantic Maps
A novel zero-shot Signal Temporal Logic (STL) planning solver has been introduced by researchers for variable-map settings, capable of producing viable trajectories without the need for retraining. This approach combines a map-conditioned Transformer architecture with a streamlined heuristic to effectively manage intricate disjunctive (OR) subformulas. To maintain temporal grounding and logical consistency across divided sub-tasks, Transitive Reinforcement Learning (TRL) is employed. Testing on dynamic semantic maps featuring various obstacle configurations shows consistent improvements, underscoring enhanced zero-shot generalization to evolving environments and extensive STL applicability.
Key facts
- Zero-shot STL planning solver for variable-map environments
- Integrates map-conditioned Transformer with lightweight heuristic
- Handles complex disjunctive (OR) subformulas
- Uses Transitive Reinforcement Learning (TRL) for temporal grounding
- Experiments on dynamic semantic maps with diverse obstacle layouts
- Demonstrates consistent gains and superior zero-shot generalization
Entities
Institutions
- arXiv