Ryan Hoque

I am a second year PhD student in the UC Berkeley EECS department studying Robotics and Artificial Intelligence. I am advised by Prof. Ken Goldberg and am part of the Berkeley Artificial Intelligence Research (BAIR) lab.

Before starting my PhD program, I spent some time in industry working on autonomous driving at Uber's Advanced Technologies Group (acquired by Aurora) and earned my B.S. and M.S. summa cum laude in Electrical Engineering and Computer Science, also from UC Berkeley (Go Bears!). Outside of research, I enjoy an eclectic mix of hobbies including freestyle rap, playing the piano, reading and writing about philosophy, chess, hiking, and parkour.

Email  /  CV  /  Google Scholar  /  Github  /  Writing  /  LinkedIn  /  Twitter

I'm broadly interested in developing imitation and reinforcement learning algorithms that can be applied to challenging real-world robotic manipulation tasks such as control of highly deformable structures. In addition to selected publications below, check out:

• My talk at CoRL 2021 in London on our recent work in interactive imitation learning.
• My M.S. thesis: "Robotic Fabric Manipulation with Deep Imitation Learning and Reinforcement Learning in Simulation."
• A BAIR Blog post in collaboration with Pieter Abbeel's lab summarizing our research on deformables up to May 2020.
• Issue #238 of the Import AI newsletter, read by more than 10,000 AI experts, featuring our work at the top.

An interactive imitation learning algorithm that reasons about both state novelty and risk to actively query for human interventions. The algorithm balances supervisor burden and task performance more successfully than prior robot-gated algorithms and is competitive with an oracle human-gated baseline.

We propose "context switching" between human and robot control as a measure of supervisor burden in interactive imitation learning, and propose an algorithm that significantly reduces context switching during training and execution.

Journal paper extending the VSF conference paper below with new datasets, visual dynamics models, cost functions, optimizers, and physical experiments.

By applying a model-based reinforcement learning technique we call VisuoSpatial Foresight, we train a visual dynamics model and use it to sequentially manipulate fabric toward a variety of goal images, entirely from random interaction RGBD data in simulation. Currently my most cited paper!

In this paper we learn to smooth real fabric from arbitrarily complex starting states by leveraging our access to ground truth state in simulation. As a preprint in 2019, this was one of the first papers to apply deep learning to robotic fabric manipulation, a research topic which has gained significant recent interest.

Various other things you may find interesting.

Website template from BAIR alum Jon Barron. This template seems to be quite popular!