Topic · DiggingBeagle record
State-aware interlocks for physical AI agent actions
Physically consequential agent commands need checks against current geometry, sensors and world state outside the model; a valid instruction or simulated plan does not itself establish safe execution.
- Topic kind
- pattern
Definition & limits
When an agent drives equipment, the authority to propose code is different from the capability to actuate a particular state transition. A state-aware gate computes whether an intended action is geometrically possible and allowed under measured current conditions, then permits or rejects it before execution. Google's VERID research uses a stateful digital twin to evaluate laboratory procedures against robot geometry, reagent and tip status, and reported sensor updates. In a controlled example it blocked an incompatible multi-channel pipetting maneuver before actuation. This is an interlock demonstration, not a general certification of wet-lab autonomy.
SafeDriveVLA addresses a related but separate setting: navigation-language requests can conflict with traffic and physical conditions. The paper reports improved selected simulation scores and fewer benchmark collisions from a conflict-aware control policy. Its world model is still fallible, and simulated counts do not estimate real crash risk. Unlike a cryptographic execution manifest, an embodied state gate must also address uncertain or stale sensor readings, unmodeled physics and the difference between mechanical feasibility and scientific or social legitimacy.
Controls should report both unsafe actions correctly blocked and safe actions wrongly rejected under distribution shifts. Human approval cannot substitute for mechanically impossible-operation detection; conversely a digital twin cannot authorize harmful experiments just because a movement is feasible. This is a narrower embodied-control pattern alongside the general Agent authority Concept.
Examples
- VERID rejects an eight-channel pipette command incompatible with a narrow trough.
- SafeDriveVLA selects a cautious/fallback maneuver instead of following an unsafe simulated driving instruction literally.
- A stale digital-twin state declares a step feasible although the real robot deck has changed.