Brian Yang

I am a fourth-year undergraduate at UC Berkeley majoring in Computer Science. I do research involving robot learning with Professor Sergey Levine in the Berkeley AI Research Lab. Previously, I worked with Professor Kris Pister on microrobot locomotion. I also worked at Facebook AI Research with Roberto Calandra on dextrous manipulation.

Email  /  CV  /  GitHub  /  LinkedIn

profile photo
mavric MAVRIC: Morphology-Agnostic Visual Robotic Control
Brian Yang*, Dinesh Jayaraman*, Glen Berseth, Alexei Efros, Sergey Levine
RA-L 2020, ICRA 2020 (under review)
preprint / project page / code (coming soon)

We propose MAVRIC, an approach for visumotor control that requires minimal prior knowledge of the robot's morphology and works using less than 20 seconds of training data. We demonstrate results in a variety of settings, including shaky handheld cameras, amputated arms, and previously unseen tools.

replab REPLAB: A Reproducible Low-Cost Arm Benchmark Platform for Robotic Learning
Brian Yang, Dinesh Jayaraman, Sergey Levine
ICRA 2019
arxiv (extended technical report) / project page / code

We present a reproducible, low-cost, self-contained hardware stack for benchmarking vision-based manipulation tasks.

morphology Data-efficient Learning of Morphology and Controller for a Microrobot
Thomas Liao, Grant Wang, Brian Yang, Rene Lee, Kris Pister, Sergey Levine, Roberto Calandra
ICRA 2019
arxiv / project page / code

Robot design is often a slow and difficult process requiring the iterative construction and testing of prototypes. We propose HPC-BBO, a novel algorithm to efficiently and automatically design hardware configurations and controllers for locomotion.

prototype Learning Flexible and Reusable Locomotion Primitives for a Microrobot
Brian Yang, Grant Wang, Roberto Calandra, Daniel Contreras, Sergey Levine, Kris Pister
RA-L + ICRA 2018
arxiv / project page / code

We validate Bayesian optimization as a data-efficient method to design gaits for legged microrobots. We also propose a novel approach to learning motor primitives by converting a contextual policy search problem into a multi-objective optimization task.

Template taken from here.