Lawrence Yunliang Chen

I am a researcher at OpenAI. I am broadly interested in AI and robotics.

I completely my PhD at UC Berkeley, advised by Professor Ken Goldberg, where I was a member of the AUTOLAB, which is part of the Berkeley Artificial Intelligence Research (BAIR) Lab. I was supported by the NSF Graduate Research Fellowship, and I have also been fortunate to intern at NVIDIA. Prior to that, I did my undergraduate at UCLA with double majors in Applied Math and Statistics, where I was advised by Professor Jungseock Joo.

Email  /  CV  /  Google Scholar  /  LinkedIn  /  Twitter

I'm broadly interested in robot learning, manipulation, and perception. My PhD research focuses on improving the robustness and generalization of robot policies through data and model scaling, data augmentation, and leveraging simulation data. Here are some of my research papers.

A systematic study of scaling robot augmentation and a large open-source dataset that augments OXE with 9× more robot embodiments across 16 datasets, comprising over 4.4M trajectories.

A simple recipe for co-training simulation and real data for robot manipulation.

A cloud-based toolkit that optimizes robot data storage and loading using EBML/MKV, achieving up to 70x compression and 50x faster loading with minimal impact on training performance.

A robot policy that learns new tasks by prompting with robot trajectories and without any fine-tuning.

A data augmentation pipeline that uses diffusion models to generate novel robots and camera viewpoints.

Zero-shot transfer a visuomotor policy trained on one robot to unseen robot embodiments by cross-painting the images.

An open-source generalist robot policy trained on a mixture of 25 datasets from the Open X-Embodiment dataset.

A dataset that contains 76k demonstration trajectories or 350 hours of interaction data, collected across 564 scenes and 84 tasks by 50 data collectors over the course of 12 months.

We introduce the Open X-Embodiment Dataset, the largest open-source real robot dataset to date. We train two models on the robotics data mixture: RT-1-X and RT-2-X.

Given a natural language query, uses LERF based on CLIP and DINO features to perform zero-shot semantic grasping of object parts.

Uses VLMs and LLMs to create semantic distributions that can be integrated into downstream mechanical search policies.

An algorithm for grasping a single layer of plastic bags and fabrics using purely visual feedback, and a much-improved bagging algorithm.

A semantic representation of plastic bags and an algorithm for a bimanual robot to open a plastic bag from unstructured configurations, to insert object(s) into it, and then to lift the bag.

A formalism, several new algorithms, and a benchmark for interactive fleet learning: interactive learning with multiple robots and multiple humans.

Given a new garment, quickly learn a fling action that can effectively smooth the garment with only one robot arm.

Optimize the arrangement of objects on a shelf to make them easier to retrieve and search.

Using Real2Sim2Real to learn to manipulate a free-end cable so that the endpoint can accurately reach desired target positions on a planar surface.

A new suction cup design that contains multiple chambers for gripping and haptic exploration.

Analysis of annotation bias for many large public facial expression recognition datasets.

Berkeley EECS C106A/206A (Introduction to Robotics) Final Project

Website template from Jon Barron