In this talk I will briefly review of some of the work done in the area of terrain adaptation with autonomous agents, and propose a dissertation topic that deals with the question: How can an autonomous agent learn methods for adapting its behaviors to arbitrary new terrains? Works in terrain adaptation are typically some combination of two main approaches. Several of the works employ purely reactive approaches, in which posture and balance controls are used to stabilize a robot. Of the remaining works, many focus on motion planning, in which all possible motions for a robot are considered based on some representation the robot has of the environment it is in. The former approach, though it is adaptive, is limited in that it can only generalize to a range of terrains that it was intended for. The latter approach, on the other hand, usually lacks the ability to react quickly to changes in the environment. In general, both methods do not learn from their experience.
In contrast to the above approaches, I want to consider the feasibility of an agent that learns about itself and its environment, and develops reactive solutions in an online manner.