lisenkaci/revolio

Revolio is a runtime that lets you control a physical robot with AI. You describe your robot in a YAML configuration file, run a single Node process, and the system connects to the robot, watches its camera, and acts on natural-language goals you type in. The core idea is a separation between the agent loop (the software that talks to an AI model, watches the camera feed, and issues commands) and the robot definition (what sensors and motors your specific robot has, and how to communicate with it). The framework code never changes. You add a new folder with a YAML manifest describing your robot's capabilities, and the system builds the right set of AI tools automatically from that description. Most robots need no custom code at all. Under the hood, Revolio uses Google's Gemini Robotics-ER model. It runs two layers in parallel. A fast reflex loop fires every 500 milliseconds, grabs a camera frame, asks Gemini what commands to send, and drives the robot. A slower planner layer fires periodically or on demand, looks at the full current situation and the stated goal, and produces a list of subtasks to guide the reflex loop. A watchdog sits on top: if the reflex loop stops responding but a motor is still running, it automatically idles all effectors to keep the robot safe. Each motor or servo is modeled as a small set of named states (for example, a drive motor can be idle, forward, backward, turn_left, or turn_right). The AI picks a state per tick, and the framework holds the motor in that state until told otherwise. For servos with angle limits, the framework tracks the estimated position and injects a hint into the AI's context so it does not keep trying to move a joint that is already at its limit. Setup requires Node 20 or later and a Google AI Studio API key. Your robot needs a WiFi connection, a WebSocket server accepting text commands, and an MJPEG camera stream. The reference robot in the repo is an Adeept RaspTank metal tank platform. A dry-run mode lets you test the full pipeline without hardware.