VASO: Formally Verifiable Self-Evolving Skills for Physical AI Agents

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

VASO is a novel framework designed for verification-guided self-evolution of LLM-generated robot skill contracts, addressing the critical challenge of ensuring trust and safety in AI-driven physical agents. While foundation models simplify skill creation, existing refinement methods offer only trace-level evidence, failing to guarantee temporal safety under untested conditions. VASO represents each skill as a semantic contract, coupling a formal interface for model checking with a planner-facing interface for executable behavior. It first filters logically inconsistent skill contracts, then verifies plans against global and local temporal specifications. Upon verification failure, VASO converts counterexample traces into textual gradients to update the reusable skill contract, crucially keeping foundation model weights frozen. Benchmarked on Clearpath Jackal and PX4 quadcopter tasks, VASO achieved 97.2% formal-specification compliance with fewer than 100 optimization samples and under 20 minutes per skill, outperforming traditional baselines like execution-feedback, prompt-optimization, and fine-tuning.

Key takeaway

For Robotics Engineers and AI Scientists developing physical AI agents, VASO offers a critical shift in skill assurance. You can now integrate formal verification directly into your LLM-driven skill evolution loops. This approach ensures generated robot skills meet temporal safety specifications, significantly reducing compliance failures from over 5% to 2.8% with minimal optimization. Consider adopting VASO to build more trustworthy and reusable robot behaviors, moving beyond trace-level evidence to provable safety guarantees.

Key insights

Formal verification feedback can directly refine LLM-generated robot skill contracts, enhancing safety and reusability.

Principles

• Skill contracts need formal and planner interfaces.
• Verification counterexamples are optimization signals.
• Early logical consistency checks prevent wasted effort.

Method

VASO generates skills, performs skill-level logical feasibility checks, then plan-level verification. Counterexample traces are converted into textual gradients to iteratively refine the semantic contract, freezing foundation model weights.

In practice

• Use VASO for verifiable robot skill evolution.
• Apply formal verification to LLM-generated plans.
• Automate proposition alignment for efficiency.

Topics

• Formal Verification
• Robot Skill Evolution
• Large Language Models
• Physical AI Agents
• Semantic Contracts
• Temporal Logic

Best for: Research Scientist, AI Scientist, Robotics Engineer, Machine Learning Engineer

Related on AIssential

• VASO: Formally Verifiable Self-Evolving Skills for Physical AI Agents — Formal verification can guide LLM-generated robot skill evolution, enhancing trust and compliance.
• CLASP: Language-Driven Robot Skill Selection and Composition using Task-Parameterized Learning — CLASP integrates TP-KMPs with VLMs for data-efficient, language-driven robot skill learning, selection, and composition.
• SkillAdaptor: Self-Adapting Skills for LLM Agents from Trajectories — SkillAdaptor enables stable, auditable, training-free skill adaptation for LLM agents via step-level failure attribution.
• ECM Contracts: Contract-Aware, Versioned, and Governable Capability Interfaces for Embodied Agents — Explicit, multi-dimensional contracts are crucial for safe, stable embodied agent software ecosystems.
• Agent Skills for Large Language Models: Architecture, Acquisition, Security, and the Path Forward — Agent skills modularize LLM capabilities, extending functionality dynamically while introducing significant security challenges.
• Formal verification for safety evaluation of autonomous vehicles: an interview with Abdelrahman Sayed Sayed — Formal verification of neural ODEs can provide critical safety guarantees for AI in autonomous systems.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.