Dexterous Manipulation for Dummies: Teleoperation, Imitation, Adaptation

Speaker: Irmak Guzey - https://irmakguzey.github.io

Location: AVŞ Seminar Room(North Campus, Department of Computer Engineering Building)

Short CV: Irmak is currently an assistant researcher in the Computational Intelligence, Vision, and Robotics Lab (CILVR) at NYU. She had her master’s degree from NYU. She was awarded a Fulbright scholarship in 2020. Her current research focus is dexterous manipulation with tactile and image information.

Abstract: Dexterous manipulation has been a complex problem due to its large action space and the multi-modality of the environment. Although there has been recent progress in introducing responsive frameworks for teleoperation, the difference between the morphology of the robot and the human hands and the lack of touch feedback for the user during teleoperation make it challenging to collect large amounts of expert demonstrations. Multi-modality and large action space make simple behavior cloning algorithms fail, and larger models that can represent the multi-modality tend to overfit when we have a low number of expert demonstrations. Consequently, over the past year, I have been exploring ways to develop dexterous policies in a multi-modal environment with limited data. My approaches so far have focused on (a) learning representations from non-task specific robotic play data and using them for downstream tasks and (b) training residual policies with online reinforcement learning (RL). In my talk, I will be talking about the approaches for collecting dexterous demonstrations and training robust & adaptable policies with this limited data.