Automated Skill Optimization via Formal Verification for Embodied Agents

Yunhao Yang, Neel P. Bhatt, Kevin Wang, Zhangyang Wang, ufuk topcu

Keywords: Skill.MD, Formal Verification, Formal Methods, Planning, Robotics

TL;DR: We develop a method that synthesizes and iteratively refines reusable robot skills with formal safety guarantees, achieving high compliance with task constraints without retraining the underlying model.

Abstract: Skill abstractions are high-level modules that enable embodied agents to compose complex behaviors from reusable components. Recent foundation models further extend this planning approach by generating skills from natural language, improving flexibility and accessibility. However, such skills lack formal guarantees, as their behaviors may violate safety or task constraints, limiting reliability in real-world deployment. We formalize verified skill synthesis, the problem of expanding a skill library while preserving global safety specifications expressed in temporal logic. Each skill consists of a formal local rule and a natural language contract, both produced by a foundation model. The contract serves as the planner-facing representation of the skill and the optimization target for improving plan quality. We introduce VASO, a closed-loop optimization process that verifies each local rule against the global specifications and refines the contract using model-checking feedback to improve specification compliance, without updating model weights. Evaluations on the Jackal ClearPath ground robot and a PX4 quadcopter show that plans guided by the optimized skills achieve 97.2% compliance with formal specifications using fewer than 100 optimization samples and under 20 minutes of optimization per skill.

Submission Number: 28