UKRI AI CDT in Dependable and Deployable AI for Robotics (CDT-D2AIR)

Sophie Higham

Project
Towards Safety and Reliability in Continually Adapting Embodied Agents

Summary
To operate effectively in dynamic, real-world environments, some form of continual adaptation will be required from autonomous agents. However, existing work in continual learning primarily focuses on task performance, while much of the embodied AI safety research assumes static, non-adaptive models. Evaluating the safety and reliability of continually learning and adapting agents therefore remains an open problem. This research therefore seeks to evaluate reliability- and safety-aware learning for embodied agents, with a focus on how safety and reliability are affected by continual adaptation in dynamic environments.