Abstract

This disclosure describes how the skills of governed robots are traced, admitted, restricted and revoked. A skill is any loadable capability that drives a robot's actions, such as a learned policy, a trained model, a motion program or a task script. Every skill version carries a lineage record, signed by the parties responsible: the skill or base model it was derived from, the sources of the data used to train or tune it, the maker and each party that modified it, the tests it passed and the robot models and places for which it is certified. Lineage records link into a tree, so that any skill can be traced back to its origins and forward to all skills derived from it. A robot's safety plane loads a skill only if its lineage verifies unbroken to an accredited root and its certification covers this robot model, hardware revision and place. A skill derived from another inherits the restrictions of its parent unless and until the derived skill earns its own certification. When a defect is found in any skill, such as a hazardous behaviour, a tampered dataset or a failed test, the finding propagates down the lineage tree, and every descendant skill on every robot in the field is restricted or revoked according to declared rules, with each robot stepping down the affected class of action. The field record of each skill version, including rehearsal confidence, incidents and near misses, is pooled across all robots that run it, so that a weakness seen on a few robots informs all. Every load, refusal, restriction, revocation and pooled finding is a signed record.

Creative Commons License

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

Share

COinS