Luna Robotics

Robots should be taught, not programmed.

General-purpose manipulation is one of the hardest engineering problems of our time.

We have started Luna Robotics, an applied robotics lab with one focus: robots that learn to act in the physical world.

Robots today are programmed. Every motion is scripted and every environment is controlled. That approach has carried industrial automation a long way and it has run out of road, because the tasks people actually want a robot to do are defined by variation, and you cannot script variation.

We build learned manipulation end to end: teleoperation systems that collect high-quality demonstrations, vision-language-action policies trained on that data, and a runtime fast enough to execute them on real hardware.

We work across the whole stack because the failures that matter live between the parts. Between data collection and training. Between simulation and hardware. Between a benchmark score and a robot that is actually useful.

The research is moving quickly. The gap is engineering — the distance between a policy that works in a published demo and one that works in an unfamiliar room, without a researcher standing next to it. Closing that gap is the whole job.

We are early, and we are building the first system now.