New Framework Merges AI Language Models with Symbolic Planning for Multi-Robot Coordination

Bridging natural language and formal reasoning, a new research framework enables teams of robots to understand complex task descriptions and execute coordinated plans with human-interpretable logic.

How PLANTOR Works#

The approach, called PLANTOR, accepts task instructions in plain English and transforms them into executable multi-robot strategies. The system uses large language models to generate a formal Prolog knowledge-base—a structured representation of the problem domain. Rather than trusting the LLM output blindly, PLANTOR applies automated consistency checks to spot and correct modeling mistakes, then passes the validated knowledge-base to a symbolic planner.

From there, the pipeline unfolds in stages: a high-level symbolic plan emerges first, which is then refined into concrete robot actions, followed by a temporally optimized schedule converted into a behavior tree that robots can execute. The design philosophy emphasizes transparency—the generated knowledge-base, intermediate plans, and scheduling constraints remain visible and inspectable throughout.

Evaluation and Real-World Testing#

The researchers evaluated their approach on benchmark scenarios adapted from Blocks World and Grippers puzzles, testing multiple language models to assess both the quality of generated knowledge-bases and the speed of the planning pipeline. They also demonstrated end-to-end execution on an actual multi-arm assembly task, showing the system works beyond simulation.

Key Findings and Implications#

Results reveal that LLM-generated knowledge-bases can substantially cut the manual modeling burden—historically one of the most labor-intensive steps in robotics planning. However, the output still sometimes requires human review and correction. The authors advocate for a hybrid workflow: deploy language models to produce the structured symbolic blueprints, but keep correctness-critical planning and scheduling logic fully symbolic and auditable. This approach balances the convenience of natural-language input with the safety guarantees that roboticists and engineers need.