Research > Robot Learning

Inverse Reinforcement Learning

Apprenticeship learning is an emerging learning paradigm in robotics, often utilized in learning from demonstration(LfD) or in imitation learning. Specifically, apprenticeship learning focuses on learning the essence behind the method demonstrated by the expert to better generalize it to similar tasks and domains. Our task at hand is the sorting of good (unblemished) and bad (blemished) onions on an industrial conveyor belt. This being done by observing the expert perform the sorting and then using inverse reinforcement learning methods to learn the task. Analogous to many robotics domains, this domain also presents problems of partial observability of state-action pairs and sensory noise that comes from sensors such as cameras used for the observation. The challenge is to use the data available to learn to the reward function used by the expert and to collaborate with the sorter.