One mind. Many machines.
A single learned policy that drives different machines, starting with wheel loaders. Built on a world model trained on our own data collected on real sites. Instead of programming a controller, we trained a model on how operators do the work. It keeps getting better with every machine and site we add.
A fleet is not one machine.
It is dozens of different ones.
One automation system per machine. Its own data, its own model, its own tuning. Nothing learned on one machine ever reaches the next, and every new machine starts again from zero.
One policy, trained across machines at once. The same physical job is learned once and shared by every machine that can do it. A new machine joins the fleet as an identifier, not a new project.
Same job. Different machines.
Most of the job is the same on every machine: identifying the pile, filling the bucket, carrying the load. Operator-1 learns it once from all of them, so each machine does better than it would if trained alone, and every machine we add improves the rest.
One site. Every site.
A model trained on one site only knows that site. Operator-1 learns from all of them at once, so what works on one site carries over to the next. Every new site we add makes it better everywhere else.
Five streams in.
One decision out.
Operator-1 combines sensor data and task instructions in one model. It plans and generates actions for the machine it is driving, adjusting as conditions change.
Thinks slow. Acts fast.
A good operator works out what the site needs while keeping the machine moving smoothly. Operator-1 does both at once.
Reads the site. Decides what matters.
Understands the scene and the instruction together. Which pile to load from, which truck to fill, and whether the person at the edge of the frame changes the plan.
Controls the machine many times a second.
Turns the plan into machine commands and keeps correcting as the bucket bites, the wheels slip and the load shifts. Continues acting while the reasoning runs.
Ready for real site conditions.
Working in dust, darkness and snow.
Water on the lens, no light at all, dust so thick you can't see the bucket. The world model has seen it on real sites, so none of it is new on day one.
Working through sensor dropout.
We cut streams on purpose during training. So a blinded camera on site is nothing new. The model has handled it thousands of times.
Trained in simulation. Runs on site.
The same policy runs in simulation and in the field. In simulation it works through thousands of cycles overnight. Every mistake there costs nothing. On site, nothing changes: the same weights that ran in simulation drive the real machine.
Work on the heavy stuff.
Research, engineering and field roles in Berlin and Potsdam. Come and teach some of the world's largest machines to work alone.
