PaCo-VLA: Passivity-Shielded Compliance Prior for Contact-Rich Vision-Language-Action Manipulation

Researchers introduce PaCo-VLA, a safety framework that shields Vision-Language-Action AI models with passivity-based compliance controls for contact-rich robotic manipulation tasks. The system treats VLA outputs as proposals rather than direct commands, using high-frequency energy monitoring to prevent unsafe interactions while maintaining semantic understanding for tasks like connector insertion.

PaCo-VLA addresses a fundamental challenge in deploying large foundation models for physical robotics: VLA systems excel at semantic reasoning but operate at low frequencies unsuitable for real-time force control. The research team's innovation decouples the decision-making layer from the control layer, allowing high-level semantic understanding to coexist with rigorous safety guarantees. This architectural separation matters because contact-rich tasks—insertion, assembly, manipulation—require precise force regulation that foundation models cannot reliably provide at required control frequencies. The passivity shield acts as a runtime enforcer, using energy accounting to validate whether model proposals maintain safe physical behavior. The framework's ability to maintain zero passivity violations even under adversarial conditions demonstrates meaningful progress toward trustworthy AI deployment in safety-critical domains. Beyond robotics, this approach suggests a broader pattern: foundation models can contribute semantic insight without direct control authority, delegating physical safety to specialized subsystems. The experimental validation through both simulation and real connector-insertion tasks provides concrete evidence the method works in practice. For the AI and robotics community, this represents a practical pathway for integrating VLAs into production systems without sacrificing safety margins. The work establishes formal runtime contracts—critical for domains where failures cause physical damage or injury. Future adoption depends on whether similar approaches scale to more complex manipulation tasks and whether the overhead of real-time passivity checking remains acceptable for practical applications.

• →PaCo-VLA decouples semantic reasoning from force control by treating VLA outputs as compliance proposals rather than direct motor commands.
• →High-frequency passivity shields using energy-tank accounting prevent invalid or unsafe model predictions from affecting robot behavior.
• →The framework achieved superior precision in connector-insertion tasks while maintaining zero passivity violations under adversarial conditions.
• →This architecture enables causal evaluation that isolates semantic model contributions from geometric shortcuts or spurious correlations.
• →The approach establishes a formal runtime contract for deploying foundation models in contact-rich manipulation domains.