Learning Plannable Representations
Humans have a remarkable capability to predict and plan complex manipulations of objects. For example, we can predict what will happen if we fold a piece of paper, and we can also plan actions to fold the paper to fit into an envelope. This `understanding' of objects allows us to safely plan and execute complex manipulation strategies every day. Motivated by this observation, in this work we study planning in dynamical systems with high-dimensional observations, such as images. We propose a framework for learning low-dimensional and structured representations of observations from a dynamical system, which can be used for planning with conventional AI planning algorithms. We train our model using videos of the system taken in an unsupervised exploration mode. Then, given some goal observation, our method predicts a realizable sequence of observations that transition the system from its current position towards the goal. We demonstrate our approach on a rope manipulation domain.
Aviv Tamar is a postdoc at UC Berkeley's Artificial Intelligence Research lab, and will start as assistant professor in the Electrical Engineering department at Technion, Israel Institute for Technology. Aviv has received his PhD and MSc in 2015 and 2011, both from Technion. His research intersects reinforcement learning, representation learning, and robotics. He is the recipient of several fellowships, including the Technion Viterbi scholarship and the Clore scholarship, and his work received the 2016 NIPS best paper award.