FANUC cobot takes spoken commands, trained before the arm existed
Lukas Ziegler says the team built the policy against a digital twin because the physical hardware was not available yet.
A FANUC CRX cobot at FANUC Europe's booth at AMB took commands in plain language and picked the object it was asked for. Vision-based object recognition worked out what was in front of it, and NVIDIA's GR00T vision-language-action model (a model that turns camera images and text into robot actions) planned the grasp and the action. Lukas Ziegler, who posted the demo, says Cosmos is used instead depending on the application.
The part worth sitting with is not the pick. It is the order things happened in.
The policy came first, the arm came second
Ziegler says the team developed against NVIDIA Isaac Sim as a digital twin because the physical hardware was not available yet. That same simulation was then used to prepare the data for fine-tuning GR00T. So the simulator did double duty. It was both the training environment and the stand-in for a machine nobody could touch.
Ziegler's framing is blunt about what that means for a development schedule. Hardware availability stops gating software progress, and by the time the physical arm shows up there is already a trained policy waiting for it. Anyone who has watched a robotics team sit idle waiting on a delivery date knows why that matters.
Instruct, do not program
The pitch on the floor is a cobot you instruct rather than program. You describe the outcome and the model works out the motion. Ziegler describes the direction as robots that understand a task well enough to adapt the action to what they actually see.
Shown versus claimed
What was shown is a trade show booth demo, one arm, spoken instruction in, object picked out. Ziegler's post does not give success rates, cycle times, or how many different objects the system handled, and it does not say how much fine-tuning data came out of the simulation. Nobody has published numbers on how the sim-trained policy (trained in simulation, run on the real robot) performed against a policy trained on the real arm, so the size of the transfer gap here is not known yet.
Still, the claim is a specific and checkable one about process rather than a vague capability boast. A team pointed a simulator at a robot that did not exist, trained against it, and then ran the result on the real thing in front of a crowd. That is a different kind of milestone than another humanoid folding laundry, and it is the kind that quietly changes how integrators plan a year.
So the robot learned its task before the robot existed, and the environment that taught it also stood in for the machine itself.
What ends up on the floor is a cobot you instruct rather than program.