Google's Gemma models run small robots on-device, with two new demos

Google's Gemma team says a Reachy Mini robot car runs two Gemma models locally on the LiteRT runtime.

Gemma 4 E4B handles voice conversation and issues tool calls that start and stop the car, while EmbeddingGemma 2 runs on an Arduino VentunoQ board and steers the robot from live camera input. The team says the whole loop stays private, offline and low-latency, but it gives no latency or accuracy figures.

Earlier, Google's Paige Bailey posted on X that small on-device multimodal Gemma models have striking potential in robotics. Her post shared a demo by Pat Loeber of a robot running Gemma 4 E2B at NVIDIA GTC. That post gave no performance figures, deployment details or availability information.