noietch/eva-client

EVA-Client is a framework for working with real robots, built to handle deployment, evaluation, and data collection all from a browser based console. The README describes it as one policy, any robot, meaning the same tool can drive different kinds of physical robot arms and humanoids through a shared workflow, letting you debug, record, evaluate, and visualize what the robot is doing without switching tools. On the deployment side, you write a single Python config file that wires together a transport layer, such as ROS1, ROS2, ZeroMQ, or replaying an offline dataset, a policy backend that runs the actual AI model, such as OpenPI, StarVLA, or GR00T, and an inference strategy that decides how actions are computed and timed. The framework currently supports six real robots, ranging from dual arm setups to full humanoids, and lets you switch between them from the same debug panel. For evaluation, EVA-Client can run sweeps across multiple model checkpoints and records detailed data for every trial, including camera video, a 3D scene reconstruction, per dimension state charts, and scores tied to specific prompts, all replayed later in a results tab. For data collection, you can teleoperate a robot and capture the session directly into the LeRobot dataset format, with a background process handling saving, an in browser pass or fail review step, and video encoding, so that both human demonstrations and AI generated rollouts share the same file layout on disk. A newer addition lets a person step in during a live policy run to correct or guide it, with the resulting interventions saved as extra training data, and there is also a real time chart showing an estimated value or quality signal as the robot works. For advanced users, the entire deploy, evaluate, and collect loop can also be run headless from a single config file over a socket connection, without needing the browser interface at all. This is a research and robotics engineering tool aimed at people who already work with physical robot hardware, not a general audience product.