Patrick Yin

I am a PhD student at the University of Washington advised by Professor Abhishek Gupta. My research is graciously supported by the NSF Graduate Research Fellowship. I'm work on embodied AI, with a focus on unlocking superhuman dexterity and robustness in robotics via large-scale RL.

I am currently a part-time student researcher at Microsoft Research. Previously, I did my undergrad at UC Berkeley where I worked with Professors Sergey Levine and Kuan Fang in the Berkeley Artificial Intelligence Research (BAIR) Lab.

Email  |  CV  |  Scholar  |  Twitter  |  Github  |  CV of Failures  |  LinkedIn

We leverage large-scale RL and diverse resets to solve dexterous, contact-rich manipulation tasks with no reward engineering or demos. We distill to RGB and show zero-shot sim2real transfer.

We propose Simulation-Guided Fine-Tuning (SGFT) - a simple, general sim2real framework which extracts structured exploration priors from simulation to accelerate real world RL.

A large, diverse robot manipulation dataset with 76k demonstration trajectories.

We propose a learning system that can leverage a small amount of real-world data to autonomously refine a simulation model, enabling sim-to-real transfer for real-world robotic manipulation tasks.

We discover that a shallow and wide architecture can boost the performance of contrastive RL approaches on simulated benchmarks. Additionally, we demonstrate that contrastive approaches can solve real-world robotic manipulation tasks.

A large, open-source real robot dataset with 1M+ real robot trajectories spanning 22 robot embodiments, from single robot arms to bi-manual robots and quadrupeds.

We propose Fine-Tuning with Lossy Affordance Planner (FLAP), a framework that leverages diverse offline data for learning representations, goal-conditioned policies, and affordance models that enable rapid fine-tuning to new tasks in target scenes.

We propose Planning to Practice (PTP), a method which makes it practical to train goal-conditioned policies for long-horizon tasks.

We propose a new form of state abstraction called goal-conditioned bisimulation that captures functional equivariance, allowing for the reuse of skills to achieve new goals in goal-conditioned reinforcement learning.