← Blog

November 2026

What LeRobot got right

Before LeRobot, getting started with robot learning required you to already know robot learning. The ecosystem was fragmented across research lab repos, each with its own conventions, its own data formats, and its own assumptions about what hardware you had. If you were coming from a machine learning background and wanted to train a policy, you were looking at weeks of setup before you could run your first experiment.

LeRobot changed that. It gave the community a shared baseline: a common dataset format, a small set of well maintained policy implementations, and hardware support for affordable robots that you could actually buy and run at home. For the first time, a researcher at a university without a robotics lab budget could collect demonstrations, train a policy, and evaluate it on a real robot in a weekend. That is a genuinely significant thing.

What it got right at a deeper level was the emphasis on standardization. The LeRobot dataset format is not perfect but it is shared. When a team publishes a dataset in that format, every other team can load it without writing custom parsing code. When a new policy architecture is added to the library, it trains on the same data as every existing one. Comparisons become meaningful in a way they were not before.

The project also made an important philosophical bet: that the best way to advance robot learning was to lower the floor, not just raise the ceiling. A lot of robotics research is focused on state of the art results on expensive hardware in controlled environments. LeRobot bet that more people doing simpler experiments on cheaper hardware would produce more useful knowledge in the long run. I think that bet is paying off.

I want to be honest about what it did not solve, because the gaps are real. The dataset format captures observations and actions but does not capture failure annotations. There is no standard way to record why a policy failed on a given episode or which step in a task broke down. The visualization tools are minimal. The workflow from collecting demonstrations to evaluating a trained policy still requires you to drop into the command line and manage files manually across different directories. None of this is a criticism of the team, who built something remarkable with limited resources. It is an observation about where the work continues.

Kepler is built on the assumption that LeRobot got the hard part right. The data format, the policy implementations, the community of people who care about making robot learning accessible: these are real assets. What we want to add is the workspace layer that makes working with that ecosystem feel as good as the ecosystem deserves.

If you have spent time with LeRobot and have opinions about what a better workflow would look like, we want to talk to you. The people who built on this project understand the problems better than anyone.