Model Utility Get in touch

Physical environments for autonomous agents

The environment where agents run physical experiments and train robots, on their own.

Model Utility turns photographs of a real lab or factory into an accurate 3D environment with physics. Agents then command a robot inside it, to carry out experiments or to train humanoids and robots.

Talk to us See how it works

Fig. 1a Photographs of a real bench.
hinge 0–92°12.2 kg246 g13.6 gμ 0.5g
Fig. 1b The same bench rebuilt in 3D, with physics.
agent
Fig. 1c An agent commands a robot at that bench.

Trusted by, via Entrepreneurs First

They back Entrepreneurs First, which Model Utility is part of.

I Vision

Agents can plan an experiment. They still have nowhere to run it.

An AI agent can read the literature, form a hypothesis and write the protocol. Then the work stops, because the next step is physical. Someone has to pick up the tube, load the machine and watch what happens.

Robots have the opposite problem. The bodies are ready, but they learn by practice, and practice on a real bench is slow, costly and easy to break. A generic simulated room does not help: it is not your lab.

We build the missing place: your real space, rebuilt in simulation, accurate enough for an agent to work in it as if it were there.

II How it works

From photographs to a working environment, in three steps.

  1. i.

    Take pictures of the space

    It starts with a camera. The lab or the factory floor is photographed as it is: the benches, the instruments, and whatever was left on them.

  2. ii.

    We rebuild it in 3D, with physics

    The pictures become an accurate 3D environment in which every object is its own thing. It has weight and friction, and its lids and doors open on their hinges.

  3. iii.

    Agents command a robot inside it

    An agent sets up a task, tells a robot what to do, and reads what happened. It can try again as often as it needs to, because nothing here can break.

Not a generic simulated room. Yours, down to the bottle someone left on the bench.

III What it is for

One environment, two kinds of work.

Autonomous experiments

An agent plans an experiment, runs it with a robot, measures the result and decides what to try next. The loop closes without a person at the bench.

Training humanoids and robots

A robot practises the real task in the real layout, as many times as it takes, before it ever touches the actual equipment.

  1. AgentSet up the bench: four tubes in the rack, centrifuge lid open.
  2. SimScene ready. Physics on.
  3. AgentPick up tube 3.
  4. RobotGrasped. 13.6 g, no slip.
  5. AgentPlace it in the rotor, then close the lid.
  6. RobotPlaced. Lid closed.
  7. AgentThat matches the plan. Next, tube 4.
Fig. 2 An agent at the bench: it sets the task, commands the robot and reads the result. Illustration.

IV Contact

Have a lab, a factory or a robot? Let's build its environment.

We are looking for labs and factories that want their space rebuilt as an environment, and for teams building agents, humanoids and robots that need somewhere real to work.

contact@modelutility.com