StandardRL
USB-C, USB do.
Fast, real-world reinforcement learning for any computer.
The device
A real-world endpoint for computer-use agents.
StandardRL is a small device that sits between an agent and an ordinary computer. Plug it in and the research boundary becomes physical: it reads the screen, decides, and types and clicks back — all through interfaces the host already trusts. No drivers, no agent API, no instrumentation on the machine being used.
- 01Pixels inThe host's display is captured as the observation.
- 02PolicyA small model runs on the device; larger models run on the server.
- 03HID outThe decision returns as ordinary keyboard and mouse input.
Observe
The screen is the state.
There is no special telemetry — only what a person would see. StandardRL captures the display exactly as it is drawn, so a policy trained against rendered pixels in simulation meets the same kind of observation in the real world. A compact encoder runs on the device itself, close to the host.
Act
Actions are ordinary input.
The host sees a keyboard and a mouse — not a privileged automation channel. Every action the agent takes is an event any program already understands, which is exactly what makes the device work on an unmodified computer, and exactly what keeps its behaviour honest and inspectable.
Power, vision, and action — down one wire.
A single USB-C connection carries everything the device needs: power delivery, the display signal it observes, a management link, and the input it sends back. One cable in, the whole loop closed.
See the device
Put the device in your space with AR on your iPhone or iPad.
Loading StandardRL device