Google's Gemma 4 E2B Runs Locally on Robot Duck

Google For Developers recently came up with a practical demonstration of their multimodal AI model, Gemma 4 E2B, running entirely on-device on two robotic ducks built by independent maker Antoine Pirrone. The demo is called “Gemma Playground: Robot Duck” as the goal of it is to exhibit the capabilities of Gemma with the help of the robot ducks as a demonstration platform. The speaker, Xavier Plantaz, Partner Solutions Engineer at Google, showcased the two robotic ducks: one running on a Raspberry Pi 5, and the other on a Jetson Orin Nano - both processing Gemma 4 E2B multimodal AI inference locally.

The robot's on-device tech stack is very interesting. It uses Parakeet for speech-to-text, Gemma 4 E2B for AI inference, and Kokoro for text-to-speech, enabling the duck to listen, look, and respond to real-time conversations. The duck's expressive design includes LED lights, moving antennas and an "attention mode" for live interaction. The project is fully open-source under Apache 2.0, with the full build guide, code, and 3D models available on the official GitHub repo of Open Duck Mini by Antoine Pirrone. So far, a lot of makers have made their own versions of Open Duck Mini. All the builds and demonstrations can be seen on the official GitHub mentioned above.