Learning Representations of States and Actions

Brian Sallans

When placed in a new environment, an autonomous agent must sift through a a huge amount of information, and decide what is important and what is not important to its survival. Unsupervised learning techniques can help it to filter information by discovering statistical features of incoming sensory data. Reinforcement learning techniques can be used to decide which features are important for maximizing reward received from the environment. I will discuss one way in which unsupervised learning can be conbined with reinforcement learning to simultaneously discover useful features of a new environment and learn to act on these features to increase reward.

SlugBot: Towards True Robot Autonomy

Ian Kelly

Two key aspects of most living things is their ability to exploit natural sources of energy within their environment and their ability to carry out appropriate behaviour in a range of different conditions. In this talk I will describe the progress of a current project that is attempting to develop a robot with an equivalent capability - a robot capable of autonomous action on agricultural land. The robot will sustain itself by hunting and catching slugs. Consequently the biomass will be fermented to produce biogas, which will be converted to electricity by a fuel cell to provide the robot's energy.