UMass Machine Learning and Friends Lunch | Main / Sensorimotor Abstraction Selection For Autonomous Robot Skill Acquisition

One of the major obstacles to autonomous robot skill acquisition is that most useful robots have rich sensors, and learning using all features extracted from all sensors is immediately infeasible. A common solution is to use a much smaller set of features hand-selected to be task relevant. While there do exist algorithms for performing this feature selection (or abstraction) automatically, none are appropriate for real-time learning. I will present an intermediate solution to this problem where the robot has a set of candidate abstractions, and selects one to use when learning a new skill. Although this still requires abstraction design it shifts the designer's role out of the skill acquisition loop and thus removes a barrier to autonomy. I will present an algorithm for state space selection and show that it selects an appropriate state space for skill acquisition in a sequence of tasks, given an initial sample trajectory.