Skip to content

Latest commit

 

History

History
167 lines (128 loc) · 5.48 KB

index.md

File metadata and controls

167 lines (128 loc) · 5.48 KB

Unitree G1 — Getting Started

The Unitree G1 is a humanoid robot platform with full-body locomotion, arm gesture control, and agentic capabilities — no ROS required for basic operation.

Requirements

  • Unitree G1 (stock firmware)
  • Ubuntu 22.04/24.04 with CUDA GPU (recommended), or macOS (experimental)
  • Python 3.12
  • ZED camera (mounted at chest height) for perception blueprints
  • ROS 2 for navigation (the G1 navigation stack uses ROS nav)

Install

First, install system dependencies for your platform:

Then install DimOS:

uv venv --python "3.12"
source .venv/bin/activate
uv pip install 'dimos[base,unitree]'

MuJoCo Simulation

No hardware? Start with simulation:

uv pip install 'dimos[base,unitree,sim]'
dimos --simulation run unitree-g1-basic-sim

This runs the G1 in MuJoCo with the native A* navigation stack — same blueprint structure, simulated robot. Opens the command center at localhost:7779 with Rerun 3D visualization.

Run on Your G1

export ROBOT_IP=<YOUR_G1_IP>
dimos run unitree-g1-basic

DimOS connects via WebRTC, starts the ROS navigation stack, and opens the command center.

What's Running

Module What It Does
G1Connection WebRTC connection to the robot — streams video, odometry
Webcam ZED camera capture (stereo left, 15 fps)
VoxelGridMapper Builds a 3D voxel map using column-carving (CUDA accelerated)
CostMapper Converts 3D map → 2D costmap via terrain slope analysis
WavefrontFrontierExplorer Autonomous exploration of unmapped areas
ROSNav ROS 2 navigation integration for path planning
RerunBridge 3D visualization in browser
WebsocketVis Command center at localhost:7779

Send Goals

From the command center (localhost:7779):

  • Click on the map to set navigation goals
  • Toggle autonomous exploration
  • Monitor robot pose, costmap, and planned path

Agentic Control

Natural language control with an LLM agent that understands physical space and can command arm gestures:

export OPENAI_API_KEY=<YOUR_KEY>
export ROBOT_IP=<YOUR_G1_IP>
dimos run unitree-g1-agentic

Then use the human CLI:

humancli
> wave hello
> explore the room
> give me a high five

The agent subscribes to camera and spatial memory streams and has access to G1-specific skills including arm gestures and movement modes.

Arm Gestures

The G1 agent can perform expressive arm gestures:

Gesture Description
Handshake Perform a handshake gesture with the right hand
HighFive Give a high five with the right hand
Hug Perform a hugging gesture with both arms
HighWave Wave with the hand raised high
Clap Clap hands together
FaceWave Wave near the face level
LeftKiss Blow a kiss with the left hand
ArmHeart Make a heart shape with both arms overhead
RightHeart Make a heart gesture with the right hand
HandsUp Raise both hands up in the air
RightHandUp Raise only the right hand up
Reject Make a rejection or "no" gesture
CancelAction Cancel any current arm action and return to neutral

Movement Modes

Mode Description
WalkMode Normal walking
WalkControlWaist Walking with waist control
RunMode Running

Keyboard Teleop

Direct keyboard control via a pygame-based joystick:

export ROBOT_IP=<YOUR_G1_IP>
dimos run unitree-g1-joystick

Available Blueprints

Blueprint Description
unitree-g1-basic Connection + ROS navigation + visualization
unitree-g1-basic-sim Simulation with A* navigation
unitree-g1 Navigation + perception + spatial memory
unitree-g1-sim Simulation with perception + spatial memory
unitree-g1-agentic Full stack with LLM agent and G1 skills
unitree-g1-agentic-sim Agentic stack in simulation
unitree-g1-full Agentic + SHM image transport + keyboard teleop
unitree-g1-joystick Navigation + keyboard teleop
unitree-g1-detection Navigation + YOLO person detection and tracking
unitree-g1-shm Navigation + perception with shared memory image transport
uintree-g1-primitive-no-nav Sensors + visualization only (no navigation, base for custom blueprints)

Blueprint Hierarchy

Blueprints compose incrementally:

primitive (sensors + vis)
├── basic (+ connection + navigation)
│   ├── basic-sim (sim connection + A* nav)
│   ├── joystick (+ keyboard teleop)
│   └── detection (+ YOLO person tracking)
├── perceptive (+ spatial memory + object tracking)
│   ├── sim (sim variant)
│   └── shm (+ shared memory transport)
└── agentic (+ LLM agent + G1 skills)
    ├── agentic-sim (sim variant)
    └── full (+ SHM + keyboard teleop)

Deep Dive