EvoPlan framework gives robots safe plans with evolutionary neuro-symbolic AI
New pipeline mines safety rules from driving logs for guaranteed robot plans.
EvoPlan, developed by Bhavya Sai Nukapotula and colleagues, addresses the gap between LLM-based robot planners (fluent but unsafe) and classical PDDL planners (safe but rigid). The framework consists of three parts, all using a locally-hosted open-weight model for on-robot deployment. First, an offline procedure mines a global Signal Temporal Logic (STL) constraint from one-class demonstration data—such as nuPlan driving logs or SCAND teleoperation—by generating counterfactual negatives and using an LLM violation generator, then fitting the constraint via evolutionary search. This mined rule shields downstream policies (e.g., a vision-language driving policy on Bench2Drive).
Second, an evolutionary PDDL planner uses the LLM to propose and repair plans, with programmatic validators selecting survivors over iterations, achieving strong performance on the ALFWorld benchmark and remaining robust when goal vocabulary mismatches action-model vocabulary. Third, a constrained execution loop compiles the plan into waypoints, checks them against the mined STL constraint, and triggers re-planning on violation. The full pipeline was demonstrated on the Gazebo simulator, proving that robots can achieve both LLM-level adaptability and formal spatio-temporal safety guarantees without cloud reliance.
- Mines STL constraints from one-class demonstrations using counterfactual perturbations and an LLM violation generator, tested on nuPlan and SCAND.
- Evolutionary PDDL planner beats strong baselines on ALFWorld and stays robust under vocabulary mismatches.
- Full pipeline runs entirely on-robot with a local open-weight LLM—no cloud dependency required.
Why It Matters
Combines LLM adaptability with formal safety guarantees, enabling trustworthy autonomous robots in the real world.