Active learning on fallback data

Abstract

A system for refining a trained autonomous control model is disclosed. The system includes a computing device configured to execute a simulation of a trained autonomous control model for a vehicle model in a simulation environment based on a predefined dataset defining a virtual driving environment and implement a fallback layer configured to detect a failure. In response to the fallback layer detecting the failure of the trained autonomous control model under simulation, the computing device is configured to identify an event in the simulation environment corresponding to the failure of the trained autonomous control model, select additional training data from a data corpus, the additional training data is analogous to the event, and execute a training process to refine the trained autonomous control model using the additional training data such that the trained autonomous control model learns to handle the event with fewer failures.