what is jetson-pi-edge fr?

Jetson-PI-Edge is an inference engine that lets robots run vision-language-action AI models directly on their own onboard hardware, such as NVIDIA Jetson devices, rather than sending data to a distant server. Vision-language-action, or VLA, models are a type of AI that looks at camera images and a written instruction, then decides what physical action a robot should take next. The project comes from a research paper about making these models run fast enough, in real time, on low power hardware where every millisecond of delay affects how quickly a robot can react. The engine is built on top of llama.cpp, an existing project for running AI models efficiently, and adds support for a specific family of robotics models called PI0 and PI0.5. It can automatically detect which of these two model types it has been given, and includes several performance tricks aimed at speeding up robot control, such as reusing computation graphs between requests, keeping intermediate data on the graphics card instead of shuffling it back and forth, and running several steps of action generation together instead of one at a time. To use it, a robot control program connects to a small local web server that stays running continuously. That server accepts camera images and the robot's current physical state, runs one round of AI inference, and returns the action the robot should take next along with timing details, all without needing to reload the model for every single step. The project also supports being called directly from Python code, and can plug into another framework called FlashRT for running the same model without starting a separate server. Building the project uses CMake, with an option to enable CUDA support for NVIDIA hardware. Running it requires two prepared model files in the GGUF format, one for the main language and action model and one for the vision encoder, which the documentation explains how to obtain and convert from existing checkpoints. This is a research codebase paired with a published arXiv paper, aimed at people working on robotics and embedded AI who want to build or extend real-time robot control systems, and the README lists several other robot AI models the project plans to support in the future.