Physical AI
Connecting AI to the physical world.
We are building a software layer that helps frontier coding agents work with real robots and machines: understanding the system, running bounded physical experiments and turning successful work into repeatable software.
Robots still take expert engineering to deploy.
Modern coding agents can generate sophisticated software, but physical systems introduce another layer of complexity: machine interfaces, calibration, coordinate systems, sensors, limits and real-world behaviour.
The International Federation of Robotics estimates that programming and integration can account for 50–70% of the cost of a robot application.
From code to physical result.
Ohm Lab gives a capable coding agent a structured way to understand a machine, write and test control software, inspect what physically happened and iterate.
Physical execution remains behind a deterministic boundary, and successful behaviour can be frozen into repeatable software rather than requiring a language model in the production control loop.
01
Requirement
An engineering task described in plain language.
02
Coding agent
A frontier model writes and revises the control program.
03
Ohm Lab
Structured machine access, bounded execution and measured feedback.
04
Physical machine
The robot performs the work behind a deterministic boundary.
Testing the idea on real hardware.
In our first controlled pick-and-place trial, the same Claude coding agent was given the same robot, task and 90-minute time budget.
With the Ohm Lab layer, the agent completed the task. Using the raw robot interface, it did not complete the task within the time limit and, during one attempt, drove the arm into the desk.
This is an initial result from one trial. We are now repeating the benchmark across further tasks and hardware.
Work with us.
We are looking to work with robotics integrators, machine builders and automation teams where programming, commissioning or re-tasking still consumes significant engineering time.
If you have a real machine and a well-defined engineering task, we would like to test it.