Part 1 - If it looks like a duck, and walks like a duck...
Jensen Huang stood on stage and told a clutch of little robot 'ducks' to fall over. It was the sensation of GTC Berlin, Nvidias European developer conference. The last stop on a European tour.
https://www.youtube.com/embed/v-NGxyPjcrE?start=4764&end=5518
The cuteness kind of wasn't the point. The robots came from Pollen Robotics, the Bordeaux-based robotics company that Hugging Face bought in 2025. The point was building agents to run them, with kit containing RTX Spark chips, using an Nvidia harness and an open-weight mixture-of-experts LLM (Nemotron 3.5 Bolt 74B-A9B).
All that was after the main robot training run - it comes with 7 pre-trained actions in it's small neural-network. To customise that and add more actions you need to train by reinforcement learning in a robot training programme, like Isaac Lab.
- Isaac Lab 3.0, Nvidia’s open-source robot-training framework, released at GTC Berlin with the Microduck as its showcase demo.
- Newton, the physics engine inside Isaac Lab that models how the duck’s body moves and falls. Nvidia co-developed it with Google DeepMind and Disney Research.
- Isaac Sim then checks the trained moves before they go onto the real robot.
Nvidia’s demo video shows Hermes planning the obstacle course and handing asset work to another agent, Astra. Astra built the props for the training environment in FreeCAD and Blender and set up the course in Isaac Sim. Isaac Lab then trained the duck across 4,096 parallel simulations, and the resulting policy was checked on an RTX Spark before going onto the duck.
To a lay person, a cute AI robot. To folk paying attention, a deep AI stack.
I'm writing this in increments around the day job because it is deathly detailed. The detail and extended blast radius folk 'Awwww' their way past during demos.
Next time: