They sit down, stand up again and even do a forward roll. The little robot ducks on Nvidia’s stage look playful, but they also show a way to teach a robot to move: practise on a computer first, then try it in the real world.
Nvidia chief Jensen Huang presented the robots on 21 October at the GTC technology event in Berlin. They are called Microduck and are made by Pollen Robotics. Nvidia provides technology that developers can use to train machines like these.
Practise without constantly falling over
A robot learning to move must keep its balance and control its motors at the right moment, among other things. That can be practised with a digital version of the robot. In this recreated world, developers can try movements before the real machine has a go.
According to Nvidia’s explanation of the demonstration, AI programs help train Microduck with Isaac Lab, software for practising with robots. Development uses both desktop computers and remote computing power.
The idea has something in common with AI that learns from games: a digital environment provides space to practise. The methods differ, but in both cases the usefulness of what has been learned outside that practice world must be demonstrated later.
Nvidia wants to supply more than chips
The robot ducks fit into a wider ambition. Nvidia wants to provide both the computers and the software other companies use to build robots and AI applications. Manufacturers can then use existing tools to develop their own machines.
Huang also discussed AI agents: programs that can perform tasks by accessing other software. With OpenShell, Nvidia wants to set limits on what such programs are allowed to do. That includes which files they can open and which commands they can execute.
That takes a lot of computers
Developing and running AI requires substantial computing power. In Berlin, Nvidia also presented computer systems housed in containers that are assembled and tested beforehand. The company wants these to make building AI data centres easier.
Huang expects European companies to make increasing use of their own AI. That is his vision for the future. Nvidia is trying to become a major supplier, while companies such as Huawei are working on alternatives.
A forward roll does not make a working factory robot
The Microducks demonstrate learned movements, but that does not prove they understand every spoken command or can do useful work independently. Exactly how they are controlled during the presentation is not fully explained. With larger humanoid robots too, reliable everyday use is a different challenge from a demonstration.
The interesting step is the way they practise. Developers can train movements digitally first and then test them on a real robot. Whether that approach also works under changing conditions is something the machine must demonstrate away from the stage.
