A robot needs to adapt when its environment, task or physical behaviour changes. Learning offers ways to acquire useful behaviours from data and experience, while control connects those behaviours to the constraints of a physical system.
CSE work has combined learning from demonstration, explanation-based learning and reinforcement learning. Observations can supply an abstract account of a behaviour, which then guides trial-and-error learning. Earlier systems applied this approach to bipedal locomotion, rescue robots and learning to use objects as tools—for example, using a hook to retrieve an object beyond the robot’s reach.
Physical platforms make these questions concrete: a method must cope with sensing errors, limited computation and the consequences of actions. Research projects connect learning to perception, planning and control so that progress can be assessed through robot behaviour.
Explore rUNSWift and the work of Francisco Cruz and Oliver Obst.
Research background
Learning robot behaviours: demonstrations, systems and selected publications records this earlier work.