SI Robot Field Learning™: Governed Learning from Operation for Robots, in which Nothing Learned in the Field Changes a Robot's Behaviour in Place, Every Learned Change Becomes a New Skill Version with Lineage that Must Pass Twin Rehearsal and Re-Enter Proving, Learning Can Never Loosen Certified Limits, Field Data Is Gathered under Owner Consent and Bystander Minimisation, Near Misses Are Weighted Ahead of Routine Success, and Regressions Are Watched and Rolled Back by the Safety Plane

Inventor(s)

• Edward D. H. Wang, SI OrganizationFollow

Abstract

This disclosure describes how governed robots improve from experience without becoming less safe or less accountable. Robots in the field gather data about their work: the situations they met, the actions they took and their outcomes. Under this method, nothing learned in the field changes a robot's behaviour in place. Every learned change, whether on one robot, across a fleet or by federated learning across owners, is packaged as a new skill version with a lineage record naming the parent skill, the data used and the learning method. The new version must pass rehearsal on the certified twins of the robots it is meant for and re-enter staged proving at a declared stage before it can be loaded. Learning can never loosen a robot's certified limits: any learned change that would raise speed, force, workspace or autonomy is refused unless separately certified. Field data is gathered only under the owner's declared consent, with personal data of people nearby minimised at the point of recording, and data from places or people who have opted out is excluded. Near misses, interventions and surprises are weighted ahead of routine successes, so that learning concentrates on what went nearly wrong. After a learned version is deployed, its field record is compared with its parent's, and the safety plane rolls back to the parent if it performs worse on declared measures. Every data collection, learning run, new version, test, deployment and rollback is a signed record.

Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 License.

Recommended Citation

Wang, Edward D. H., "SI Robot Field Learning™: Governed Learning from Operation for Robots, in which Nothing Learned in the Field Changes a Robot's Behaviour in Place, Every Learned Change Becomes a New Skill Version with Lineage that Must Pass Twin Rehearsal and Re-Enter Proving, Learning Can Never Loosen Certified Limits, Field Data Is Gathered under Owner Consent and Bystander Minimisation, Near Misses Are Weighted Ahead of Routine Success, and Regressions Are Watched and Rolled Back by the Safety Plane", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12896