Build Open and Fast
Robot-Use Agents

We build open, LLM-centric robot-use agents that are

  1. FastThey work in real time: you can talk to them mid-task, and they notice when the scene changes.
  2. General-purposeOne agent across robot bodies, scenes and tasks nobody scripted.
  3. Continually learningThey pick up new tasks and environments from a demonstration, or by practising on their own.

News

  • ORA is at CoRL 2026, in the Late-Breaking Results session.Paper
  • Introducing ORA: an open harness that connects open-weight language models to robots.Post

Projects

Open Robot-Use Agents

Open models operate robots. Zero-shot.

RoboDojo's Build Tower, solved in code mode: the head camera and the two wrists.

Solves tasks it was never trained on

Share of cases solved, one open-weight model.

Writes code

One program carries 3.6 robot actions a turn, against 1.1 with tool calls.

No VLA, no robot training

A language model, cameras and tools; every run is sealed and can be replayed.

Latest posts

All posts