What is realtime-vla-flash?

Realtime-VLA FLASH is a Python research project aimed at making AI-controlled robots respond much faster than usual. In robotics, a VLA, short for Vision-Language-Action model, is an AI system that looks at the world through cameras, understands a language instruction, and decides what physical actions a robot should take next. The problem this project addresses is that running these models is normally slow, which makes smooth, real-time robot control difficult. FLASH tackles this with a technique called speculative inference. Instead of running the full, heavy model for every single decision, a smaller and faster draft model predicts what the robot should do next, and the full model only steps in to check that prediction rather than computing everything from scratch. According to the README, this brings inference down to as fast as 7.8 milliseconds when using two camera views, which allows for more than 125 decisions per second, fast enough for continuous real-time control. The project also uses customized Triton kernels for serving the model, which the README states gives roughly a three times speedup on benchmark tasks compared to the standard approach. Setting this up is involved. The README points users to a separate openpi installation guide, requires cloning the repository with its submodules, and using the uv tool to install the Python environment. Getting a working system running means converting pretrained checkpoints into a specific weight format for Triton, then starting a policy server and a matching evaluation client for the LIBERO robotics benchmark. The project also includes scripts for training your own faster draft model from a cache of examples. This is research-grade robotics software written in Python, aimed at people already working with diffusion-based robot-control models such as pi0 who need to run them at real-world speeds. It builds on an earlier project called realtime-vla and the openpi framework, and the README links to an accompanying paper and a hosted model on Hugging Face for more background.