Gemma 4 VLA Running on Jetson

A VLA demo setup was introduced that runs Gemma 4 locally on an NVIDIA Jetson Orin Nano Super (8GB), where the model itself decides whether to look at the webcam.

The pipeline flows as Parakeet STT → Gemma 4 → webcam call (if needed) → Kokoro TTS, with the SPACE key controlling the start and stop of recording. Without keyword matching or hardcoded rules, the model directly calls the look_and_answer tool depending on the context of the question.

The following information is provided for reproduction.

• Required hardware: Jetson Orin Nano Super 8GB, Logitech C920 webcam, USB speaker, USB keyboard
• Local runtime environment: build llama.cpp with CUDA, and download Gemma 4 E2B-it Q4_K_M GGUF and mmproj files
• If memory is insufficient: add swap, kill Docker/unnecessary processes, and downgrade to Q3 quantization if needed
• Example server settings: --image-min-tokens 70, --image-max-tokens 70, -ngl 99, --flash-attn on, --jinja

Also, for text-only use, it can be tested more simply with the Jetson AI Lab's Docker image, but this path does not load the vision projector, so it is not suitable for VLA demo purposes. The full demo including the webcam must use the native llama.cpp configuration.