dimos run - on quadrupeds, humanoids, arms, and drones. Simple install, entirely Python, no ROS required.
It is agent native in a precise sense: robot capabilities are exposed as skills an LLM can call over MCP, so “Hey robot, go find the kitchen” is a supported control surface - and every skill also works from the CLI with no LLM involved.
Start with the Quickstart: fifteen minutes from install to replaying a robot session and talking to an agent, no hardware needed.
Capabilities at a glance
Navigation & mapping
Live SLAM with a column-carving voxel map, costmap planning and replanning, premap relocalization. Send goals by viewer click, Python, or natural language.
Perception
2D/3D detection (YOLO11, open-vocabulary YOLO-E), person tracking, fiducial markers to TF, VLM scene queries.
Agents
An LLM loop over MCP-exposed skills. Tag places, navigate by description, add your own skills in a few lines.
Memory
Three systems: a live place map for language navigation, a recorded observation store with semantic search, and experimental event memory.
Manipulation
Drake-backed motion planning, teleop (teach mode, VR, keyboard), and a collect-to-LeRobot imitation learning loop.
Platforms
Unitree Go2 and G1, Galaxea A1Z, A-750, OpenArm, xArm, Piper, drones - real, simulated, or replayed.
Start here
Quickstart
Install DimOS, replay a session, drive an agent - no hardware required.
How DimOS fits together
The mental model: modules, streams, blueprints, and how you build applications on them.