NVIDIA showcases Microduck AI robot training at GTC Berlin

At the GTC Berlin keynote, NVIDIA CEO Jensen Huang demonstrated the Microduck robot, a 25 cm bipedal machine weighing less than 1 kg. The demonstration featured four Microduck robots performing actions such as sitting, standing, and napping on cue. While promotional materials suggested the AI agents brought the robot from simulation to the real world, NVIDIA’s technical documentation clarifies that the agents' role was focused on planning an obstacle course and assembling props within the Isaac Sim environment rather than solving the sim-to-real gap.

The Microduck project is a collaboration between Hugging Face and Pollen Robotics, designed as an open platform for students and developers to research robotic motor skills. The workflow utilizes the RTX Spark platform, employing the Hermes Agent and NVIDIA’s Nemotron 3.5 Bolt model to convert prompts into step-by-step plans. These plans are then used to build assets in FreeCAD and Blender before being simulated in Isaac Sim.

Following the announcement, Hugging Face CEO Clem Delangue reported that preorders for the US$399 robots have exceeded 20,000 units.