VASO: Self-evolving robot skills with 97.2% formal verification compliance

VASO, developed by a team including Yunhao Yang and Ufuk Topcu, tackles a core problem in embodied AI: how to trust LLM-generated robot skills beyond the specific scenarios they were tested on. Existing skill-evolution loops rely on traces — showing a skill worked on sampled executions — but VASO closes the loop with formal verification. Each skill is represented as a semantic contract with two interfaces: a formal interface mapping robot states, observations, and commands to logical propositions for model checking, and a planner-facing interface that guides executable behavior generation. A model checker first filters logically inconsistent contracts, then verifies induced plans against global and local temporal safety specifications. When verification fails, VASO translates the counterexample trace into a textual gradient that updates the reusable skill contract — all while keeping foundation-model weights frozen.

On Clearpath Jackal ground robots and PX4 quadcopter tasks, VASO reached 97.2% formal-specification compliance using fewer than 100 optimization samples, significantly outperforming execution-feedback, prompt-optimization, and fine-tuning baselines. The framework is, to the authors' knowledge, the first to use formal counterexamples as optimization feedback for reusable robot skill contracts, rather than merely verifying one-off plans or tuning prompts. This means robot skills can evolve and become provably safer without modifying the underlying LLM — a major step toward trustworthy physical AI agents in open-ended environments.