My background is in software, data, and production ML. I'm now focusing on robotics, especially data, evaluation, and machine learning. How do we collect useful robot data, and how do we check whether a policy improves?
I learn by implementing recent papers and trying things on a robot. It helps me understand how the ideas work and where I still have questions.
| What I'm exploring | What I've been working on |
|---|---|
| Learning from papers | Faster-WAM, now merged into LeRobot. QF3 is an open proposal for a state-only flow actor. |
| Robotics data | All LeRobot pull requests around streaming and repeatability. |
| Robot control | SO101 control experiments and a pen holder for the arm. |
I write about robotics, reinforcement learning, and experiments on my blog. A recent article looks at world action models and VLAs.
Breaking Bellman is an experiment in learning faster with AI through interaction. You make guesses, test them in small reinforcement learning games, and see what happens. AI help is optional. I'm exploring whether this makes the concepts easier to understand. Source.
Away from robotics, Culture Calendar brings together cultural events in Austin. Source.