Deepinder Mann

Hello! I'm Deepinder, a second-year student and undergraduate researcher at UC Berkeley, studying Electrical Engineering & Computer Science (EECS). I am advised by Sergey Levine at the Robotic AI & Learning Lab @ BAIR. Previously, I have also researched computer-use agents at Apple.

Email  /  Scholar  /  Twitter  /  Github

Research

Broadly, I am interested in improving the scalability of RL algorithms, via paradigms like offline RL, unsupervised RL, and RL pre-training. My long-horizon goal is to develop large RL models that can learn in a completely unsupervised manner from the vast collections of unlabeled data on the Internet.

My current focus is on using offline goal-conditioned RL to accomplish the above.

To combat exogenous noise in the goal-conditioned RL setting, we introduce dual goal representations: representing a goal state purely in relation to other states.

We empirically show that the poor scalability of TD-learning is rooted in the so-called "curse of horizon" and suggest practical horizon reduction techniques to alleviate this problem.