Abstract
This disclosure describes a runtime regulator for machines whose high-level behaviour is chosen by an AI model (robots, drones, vehicles, industrial arms). The regulator divides the machine's control into a planning layer, which chooses goals and actions, and a stabilizing layer, which keeps the body physically safe (balance, attitude, braking, thermal and power limits). When the regulator detects that the machine is leaving its declared operating envelope, it intervenes in graded steps that act on the planning layer only: it narrows the goal set, slows the rate of new actions, restricts the action vocabulary, requires human approval and, at the last step, pauses the planner. At no step does it slow, pause or override the stabilizing layer. The only command the regulator may send towards the stabilizing layer is a request to enter a declared safe holding state for that machine class (hover or land, controlled stop, stable stance, hold position under torque limits). The boundary between the two layers, the permitted holding states and a certification test that proves the stabilizers keep running at every intervention level are declared in advance and recorded. The result is that a misbehaving or uncertain AI planner can be throttled down to zero without the machine falling, drifting, crashing or dropping a load because the safety intervention itself disturbed its low-level control.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Wang, Edward D. H., "SI Regulator™: Runtime Intervention on AI Controlled Machines that Throttles Planning and Never Stabilizing Control", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12136