Physical AI
The Jane Street Physical AI Lab at ReLU is investigating and experimenting with the emerging opportunities for machine learning in the field of robotics. With a special focus on robot learning, we deploy models on our in-house robot setup and carry out projects with industry partners. Our ambition is to produce research that meets the standards of leading international robotics laboratories.
Robots are increasingly learning tasks from demonstrations instead of hand-written controllers. With the newest models, the result is decided by the whole loop around the model: collecting good data, fine-tuning, testing, and running on a real robot. That loop is where we work.
We build at the frontier of robot learning, with vision-language-action (VLA) models at the center. We start from the newest open VLAs, fine-tune them on demonstrations we collect ourselves, and run them on real robots. We are also exploring simulation and RL fine-tuning. Our goal is to learn by building, from quick experiments to robust systems and research worth publishing.
We run the lab like an engineering team. Version control and pull requests, locked environments, and every experiment and test result logged. We work in weekly sprints, and we train on our own GPU hardware.
Our members are ambitious students who already know a lot, and they came to build hard things. We take on real problems with industry partners, and we aim to publish.
Projects and posts
Nothing posted from this lab yet. The first writeups will show up here.