Unitree Go2

The reference DimOS platform: navigation, agentic control, recording - real, simulated, or replayed

The Unitree Go2 is the reference DimOS platform - the most complete stack in the repo, and the one every bundled replay dataset and most tutorials use. Everything runs three ways: on real hardware over WebRTC, in MuJoCo simulation, or against recorded sessions with no robot at all.

• Setup your Dog - requirements, install, connecting to your Go2, and agentic control
• Simulation - try it with no hardware via replay or MuJoCo
• Mapping & Navigation - live nav, premap recording, and relocalization

Available Blueprints

Blueprint	Description
dimos run unitree-go2-basic	Connection + visualization (no navigation)
dimos run unitree-go2	Full navigation stack
dimos run unitree-go2-agentic	Navigation + LLM agent + MCP tool access
dimos run unitree-go2-agentic-ollama	Agent with local Ollama models
dimos run unitree-go2-spatial	Navigation + spatial memory
dimos run unitree-go2-detection	Navigation + object detection
dimos run unitree-go2-memory	Navigation + record lidar/odom/color_image to .db
dimos run unitree-go2-relocalization	Navigation + align live scans to a saved .pc2.lcm premap

Add --replay (recorded session) or --simulation (MuJoCo) to any of these to run without hardware.

Deep Dive

• Using navigation - send goals by viewer click, Python, or language
• Navigation stack - column-carving voxel mapping, costmap generation, A* planning
• Relocalization - record → dimos map global --export → replay or live deploy
• Tutorial: drive the Go2 with language - the agentic stack end to end
• Visualization - Rerun, performance tuning
• Streams - RxPY streams, backpressure, quality filtering
• Blueprints - composing modules