← Blog

November 2026

Why five tools

Ask any robot learning engineer what their workday looks like and they will describe the same ritual. Open VS Code, open a terminal for ROS 2, start Isaac Sim in a separate window, run a LeRobot script from the command line, and pipe the output into Rerun to see what actually happened. Five tools. Five contexts. One task.

Nobody designed this. It accumulated the way most engineering debt accumulates: each tool solved a real problem at the time it was built, and nobody had the mandate or the time to glue them together afterward. ROS 2 handles robot communication. Isaac Sim handles physics. LeRobot handles data collection and training. Rerun handles visualization. VS Code handles everything else. Each one is excellent at what it does. Together they are exhausting.

The cost is not just inconvenience. The real cost is what happens at the boundary between tools. You record a demonstration in one format, convert it for training in another, and then try to understand why the policy failed by correlating logs across three separate applications that have no shared concept of time. The failure is obvious in hindsight. Finding it costs an afternoon.

I have watched engineers spend more time managing their tools than thinking about their robots. That is not a complaint about the tools. It is an observation about what happens when a stack grows without anyone owning the full picture.

The closest analogy is software development before modern IDEs. You had editors, compilers, debuggers, and version control, all separate, all with their own workflows. The IDE did not replace any of them. It brought them into a shared context where they could talk to each other. When the compiler found an error, the editor jumped to the line. When the debugger paused on a breakpoint, the variable inspector updated automatically. The tools became aware of each other and that changed what was possible.

Physical AI is at a similar moment. The tools exist. The question is whether someone will build the context that connects them. That is what we are trying to do with Kepler. Not to replace LeRobot or Rerun or ROS 2, but to give them a shared workspace where a failure in the simulator can be traced back to the exact demonstration that caused it, in one view, without switching windows.

Five tools is not a law of nature. It is an artifact of history. We think it is time to change it.