Columbia University, NY                                                                September 2021 - Present​

​

Problem:

• Use a 5-finger robotic hand to learn to rotate an object as fast and long as possible.

• Pure reinforcement learning (RL) methods are not enough to learn. They don't learn finger-switching behavior to allow continuous rotation.

Solution:

• Tried various methods to combine human-designed controllers with RL algorithms, to improve exploration and sample efficiency.

Paper:

• Value Guided Exploration with Sub-optimal Controllers for Learning Dexterous Manipulation.

• Submitted to The 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS).