ViTL lets robots follow temporal logic commands with zero-shot navigation

Robots have long struggled with commands like 'Clean either the chair or the couch, then turn on the TV' — tasks requiring both logical disjunction and temporal ordering. Existing zero-shot navigation systems handle only single-target instructions, leaving a gap for real-world multi-step requests. To address this, researchers from Lehigh University introduce ViTL (Vision-Language Temporal Logic), a framework that combines large language models (LLMs) with formal methods. ViTL first uses an LLM to parse natural language into Linear Temporal Logic (LTL) formulas, which are then converted into Deterministic Finite Automata (DFA). These DFAs orchestrate multi-channel value maps — one per sub-task — and trigger dynamic replanning when new objects are detected, enabling the robot to follow implicit temporal and logical constraints without any task-specific training.

At the navigation level, ViTL introduces a 'directional score' that improves frontier-based exploration. Instead of computing a single value across the entire field of view, it labels candidate directions on the observation image and queries the VLM for per-direction scores. This allows the robot to make more informed decisions about which way to move. Experiments on the Habitat-Matterport 3D (HM3D) benchmark show that ViTL achieves zero-shot long-horizon task completion with temporal constraints. The directional score also boosts accuracy and efficiency on single-target navigation compared to baseline VLM-based methods. The code is available, opening the door for broader adoption in service robotics and autonomous systems.