Yuxiang Yang

I'm a Research Scientist at Google Deepmind, where my work focus on the intersection of AI and Robotics. Prior to that, I completed my PhD at University of Washington, where I work with Prof. Byron Boots at the UW Robot Learning Lab. My research interest lies in the combination of machine learning and optimal control, with an application on quadrupedal robots. Prior to PhD, I obtained my undergraduate degree at UC Berkeley, and spent two years as an AI Resident in Google.

News

• (2024/09) I defended my PhD and started my job as a research scientist in Google Deepmind.
• (2024/07) I'm so happy to get a quadrupedal robot jumping upstairs that I decided to write a news about it.
• (2024/05) We hosted the Agile Robotics workshop at ICRA 2024 and it was a full house!

Research

I'm generally interested in robotics, control, and learning. I would love to see the combination of them that solves complex, dynamic and real-world problems.

Agile Continuous Jumping in Discontinuous Terrains

By augmenting the hierarchical jumping framework (CAJun) with perception and performing careful system identification, we make the robot do continuous, terrain-adaptive jumping on challenging terrains like stairs and stepping stones.

LocoMan: Advancing Versatile Quadrupedal Dexterity with Lightweight Loco-Manipulators

We augmented the manipulation capabilities of quadrupedal robots with custom-designed lightweight grippers, and show that the resulting system can do a wide variety of dexterous manipulation tasks.

CAJun: Continuous Adaptive Jumping using a Learned Centroidal Controller

We design a GPU-accelerated, general-purpose, hierarchical framework to learn continuous, long-distance, and adaptive jumping for quadrupedal robots.

Continuous Versatile Jumping using Learned Action Residuals

We enable omni-directional jumping and jump-turns by combining heuristic controllers and learned action residuals.

Learning Semantics-Aware Locomotion Skills from Human Demonstrations

We build a framework for quadrupedal robots to learn offroad locomotion skills based on perceived terrain semantics.

Fast and Efficient Locomotion via Learned Gait Transitions

We design a hierarchical framework for quadrupedal robots to learn energy-efficient gait patterns for quadrupedal robots.

Rapidly Adaptable Legged Robots via Evolutionary Meta-Learning

We use evolutionary-strategy (ES) based meta learning to perform dynamics adaptation on a real legged robot.

ES-MAML: Simple Hessian-Free Meta Learning

We introduce ES-MAML, a new framework for solving the model agnostic meta learning (MAML) problem based on Evolution Strategies (ES).

Data Efficient Reinforcement Learning for Legged Robots

We design a model-based framework that learns to walk using less than 5 minutes of data and generalizes to unseen tasks.

NoRML: No-Reward Meta Learning

We introduce a new algorithm for meta reinforcement learning that is more effective at adapting to dynamics changes.

OpenRoACH: A Durable Open-Source Hexapedal Platform

We present a open-sourced, low-cost, ROS-enabled legged robot platform for research and education.

Projects

Python Environment for Unitree Robots

Building off the motion_imitation repo, I first developed a python-based framework for the A1 robot from Unitree. The framework includes a simulation based on Pybullet, an interface for direct sim-to-real transfer, and an reimplementation of Convex MPC Controller for basic motion control.

In 2021, I refactored the environment to remove unnecessary dependencies. The new environment is now available in the open-sourced repo of my CoRL 2021 paper.

In 2023-2024, I migrated the entire framework to Isaac Gym for GPU-accelerated simulation and training. Check here for the non-perception version, and here for the perception-integrated version.