What We're Building

Robots have been able to move boxes and weld cars for decades. What's always been expensive is teaching one a specific job. Robotics still works at a low level: ROS, motor controllers, and hand-tuned control code. There's no layer between what a robot should do and how its motors move, so every new task is custom engineering by robotics specialists. That only pays off for work repeated millions of times in exactly the same way, like a car assembly line.

Almost no real work looks like that. Most of it is varied, changes often, and happens at too small a scale to justify a purpose-built robot. Every time a task or procedure changes, someone has to reprogram the robot, so most of this work never got automated, even where a robot could physically do it.

The Bottleneck Is Deployment

Building a capable robot is no longer the hard part. Most well-funded teams can build a good prototype. The hard part is getting a robot to do a specific task reliably, in a real environment, without a large team of specialists building every deployment by hand.

Iteration is what makes this slow. In software, you can try something, see it fail, and fix it in minutes. In robotics, months of work can pass before you find out an approach doesn't work, and every real-world failure costs time, money, and sometimes the hardware itself. Without fast, cheap iteration, robotics stays expensive and only open to specialists.

Why Now

Software went through this shift first. AI agents now reason, plan, use tools, and recover from mistakes, and a large community of developers builds them every day. Almost none of that work reaches robotics, because there's nowhere for an agent to plug in. The robot speaks in joint angles and motor commands, not tasks.

Two things are changing that. Frontier models are getting good at physical tasks: they can reason about a scene, write the code to act on it, and improve with feedback. And agents can now run research on their own: an agent designs an experiment, runs it, checks the result, and decides what to try next, over and over, far faster than a person could. In robotics, that loop can take over much of the trial and error that today takes specialists months.

From Hardware to Software

We don't think one robot or one tool solves this. Getting a robot to do real work takes the whole chain: the AI that figures out the task, the simulation to test it safely, and hardware that behaves the way the simulation says it will. So we build across all of it, from the AI to the hardware.

  • AI that does the trial and error. People describe the task and judge the results. AI runs the experiments in between, so a small team, even just one person, can take on projects that used to be out of reach.
  • Simulation first. A task should be proven on a digital twin before a robot is ever deployed. Simulation is where failure is cheap, and it's where most of the iteration should happen. What the robot learns on site then makes the simulation better.
  • Open hardware. Robots and components built to be understood, repaired, and improved, so what works in simulation holds up on the real robot.

Our Products

Asimov, an Open Humanoid

Asimov is our open humanoid, about 1.2 m and 35 kg. We co-designed the hardware and the AI, so policies trained on its digital twin run on the real robot with no extra tuning. The bill of materials, CAD, and docs are public, so anyone can build, repair, and modify it, and its SDK works with the frontier model or agent of your choice. Asimov already ships to customers worldwide.

Order Asimov now, or talk to us about custom solutions, partnerships, or anything else you have in mind.

Jan, an Open-Source AI Agent

We also made Jan, an open-source AI agent for macOS, Windows, and Linux, with 7 million downloads.

Built With the Community

Community has always been central to how we work. Jan was built in the open, and Asimov's designs are public so anyone can build on them. We believe robotics will move faster bottom-up than top-down: no single lab can create every environment, train every skill, or test every hardware change, but a global community of developers, researchers, and engineers can.

The Long View

We believe dangerous, dirty, and dull work shouldn't fall to people when a robot can do it instead. That's true in a warehouse today, and one day it'll be true in places humans can't easily reach. Long-term, space.

On this page