---
title: "Unitree G1 Humanoid Robot"
space: "Factory Docs"
url: "https://factory.uga.edu/docs/equipment/unitree-g1"
updated: "2026-11-19"
---

## Overview

The Unitree G1 is a compact humanoid robotics platform designed for locomotion, manipulation, reinforcement learning, embodied AI, and human-centered robotics research. Our lab has both **G1 Air** and **G1 EDU** systems.

> **Configuration warning:** G1 hardware is highly configuration-dependent. The EDU family can range from 23 to 43 DOF depending on waist, wrist, and hand options. Record the actual joint configuration of each robot rather than treating "G1 EDU" as one fixed hardware definition.

## Air vs EDU

Unitree's current official developer documentation primarily distinguishes the basic G1 and G1 EDU hardware. "G1 Air" has also been sold as a reduced/base configuration, so use the Air's physical build and purchase configuration when an option differs from the base specification.


| Parameter                | G1 Air / Base Configuration            | G1 EDU                                      |
| ------------------------ | --------------------------------------: | -------------------------------------------: |
| Standing dimensions      | \~1320 x 450 x 200 mm                  | \~1320 x 450 x 200 mm                       |
| Folded dimensions        | \~690 x 450 x 300 mm                   | \~690 x 450 x 300 mm                        |
| Weight with battery      | \~35 kg                                | \~35 kg+ depending on configuration         |
| Total DOF                | Typically 23 on base/Air configuration | 23-43                                       |
| Leg DOF                  | 6 per leg                              | 6 per leg                                   |
| Arm DOF                  | 5 per arm                              | 5 per arm, with optional wrist DOF          |
| Waist DOF                | 1                                      | 1, optional additional 2                    |
| Maximum knee torque      | \~90 N.m base                          | Up to \~120 N.m EDU                         |
| Arm load                 | \~2 kg base                            | \~3 kg EDU                                  |
| Perception               | Depth camera + 3D LiDAR                | Depth camera + 3D LiDAR                     |
| Basic compute            | 8-core CPU                             | 8-core CPU                                  |
| High-compute module      | Configuration dependent                | Optional compute such as NVIDIA Jetson Orin |
| Secondary development    | Configuration/license dependent        | Supported                                   |
| Battery                  | 13-series Li-ion, quick release        | 13-series Li-ion, quick release             |
| Nominal battery capacity | 9000 mAh family specification          | 9000 mAh family specification               |
| Approximate endurance    | \~2 h, application dependent           | \~2 h, application dependent                |


## Kinematic Structure

The G1 is a floating-base humanoid. Unlike a fixed industrial manipulator, the robot's base pose is itself part of the dynamic state.

For the 23-DOF body:

```text
                   Head
                    |
                 Torso
              /     |     \
         Left Arm  Waist  Right Arm
           5 DOF    |       5 DOF
                    |
             +------+------+
             |             |
         Left Leg      Right Leg
           6 DOF          6 DOF

5 + 5 + 1 + 6 + 6 = 23 actuated body DOF
```

EDU configurations may add two waist DOF, wrist joints, and dexterous hands.

## Optional Dexterous Hands

The EDU can be configured with the **Dex3-1 three-finger dexterous hand**.

Each Dex3-1 has seven actively controlled finger DOF:

- thumb: 3 DOF
- index finger: 2 DOF
- middle finger: 2 DOF

Optional wrist joints and tactile sensing further change the usable manipulation model. Always make the software model match the physical robot.

## Perception

The G1 combines a head-mounted 3D LiDAR and depth camera with proprioceptive sensing from the robot body.

The lab's G1 user manual identifies a Livox MID360 LiDAR and Intel RealSense D435i depth camera on the documented configuration.

Typical embodied-AI data flow:

```text
       LiDAR          Depth Camera
          \              /
           \            /
            v          v
             Perception
                 |
                 v
          World / Robot State
                 |
          +------+------+
          |             |
          v             v
    Task Planning   Learned Policy
          |             |
          +------+------+
                 |
                 v
          Whole-Body Command
                 |
                 v
              G1 Body
```

## Development Architecture

G1 EDU development uses the Unitree SDK2 ecosystem and DDS communication. Unitree provides developer interfaces and examples for:

- basic/high-level motion
- joint motor control
- arm control
- remote-control data
- robot state
- odometry
- LiDAR
- depth camera
- SLAM/navigation
- dexterous hands
- ROS 2 communication
- reinforcement-learning control

Conceptually:

```text
Research Workstation / Onboard Compute
                 |
          unitree_sdk2 / DDS
                 |
                 v
          G1 Service Layer
                 |
          +------+------+
          |             |
          v             v
     High Motion    Joint Control
          |             |
          +------+------+
                 |
                 v
          Robot Controllers
                 |
                 v
        Motors + Sensors
```

## Develop Mode

For low-level EDU development, **Develop mode is operationally important**.

The lab's G1 manual warns that the built-in motion-control program continues transmitting commands after boot. If low-level SDK commands are sent at the same time, both command sources can conflict and cause jitter or instability.

The documented EDU workflow uses the remote-control combination `L2 + R2` to enter Develop mode while the robot is suspended and damped. Confirm behavior against the installed firmware before relying on a key combination.

The important system principle is:

```text
One active authority for a controlled joint set at a time.
```

Two independently reasonable controllers sending commands to the same actuators do not combine into a reasonable controller.

## High-Level Control

Use the high-level motion interfaces when the research problem does not require replacing Unitree's locomotion controller.

This is appropriate for:

- walking commands
- navigation
- mobile manipulation
- perception research
- task planning
- human-robot interaction
- autonomous manufacturing tasks

The robot controller remains responsible for much of the difficult dynamic stabilization problem.

## Low-Level and RL Control

Low-level control enables research into:

- reinforcement-learning locomotion
- imitation learning
- whole-body control
- custom walking policies
- sim-to-real transfer
- joint-level control
- learned manipulation

A sensible development path is:

```text
Simulation
    |
    v
Policy Validation
    |
    v
Hardware-in-the-loop / Suspended Test
    |
    v
Restricted Physical Test
    |
    v
Full Locomotion Experiment
```

Do not make unconstrained physical testing the first validation layer for a new controller.

## Networking and DDS

Unitree SDK2 uses DDS for robot/application communication and provides ROS 2 examples.

When debugging communication, isolate:

1. physical network connectivity,
2. IP/interface configuration,
3. DDS discovery,
4. topic/service availability,
5. SDK/firmware compatibility,
6. application behavior.

This distinction matters because a healthy Ethernet connection does not imply functioning DDS discovery, and functioning DDS discovery does not imply that the robot is in a mode that will accept motion commands.

## ROS 2

ROS 2 can be used as a research integration layer for perception, planning, logging, visualization, and higher-level behaviors.

```text
ROS 2 Research Stack
         |
        DDS
         |
         v
Unitree Interfaces
         |
         v
G1 Control System
```

Avoid treating a general-purpose ROS 2 node as a substitute for the robot's real-time actuator-control infrastructure unless the experiment is specifically designed around that constraint.

## Startup and Shutdown

The G1 is dynamically unstable when unpowered. A standing robot can fall when torque is removed.

The manufacturer's documented workflows therefore use either:

- supported seated startup/shutdown, or
- a protective suspension frame.

For experimental development, the suspension frame is particularly useful because it lets the robot enter powered and development states without immediately exposing the system to a full-body fall.

## Emergency / Damping Behavior

The documented remote-control procedure uses `L1 + A` to enter damping mode. In damping mode the robot may **fall or settle under gravity** rather than freezing rigidly in space.

That means an emergency action should be planned together with the physical environment. There must be somewhere safe for the robot to fall.

## Safety

The G1 is approximately 35 kg and can move at speeds approaching 2 m/s in appropriate configurations. Treat it as a powerful dynamically balancing machine, not as inherently safe because it is human-sized or research-oriented.

Major hazards include:

- robot falls
- kicks and limb impacts
- finger, knee, hip, and arm pinch points
- unexpected learned-policy behavior
- conflicting software commands
- battery depletion while standing
- payload/tool impacts
- loss of network control

Unitree's manual recommends maintaining at least \~2 m clearance from obstacles, complex terrain, crowds, and water during operation and keeping the machine in sight.

For development:

1. Use a protective suspension frame for initial low-level tests.
2. Keep people outside the fall envelope.
3. Know the damping/emergency procedure before enabling motion.
4. Validate remote control and battery state before testing.
5. Ensure only the intended controller is commanding the robot.
6. Start with reduced motion amplitude and speed.
7. Move to unsupported walking only after basic state and joint behavior are verified.

## Environment

The documented operating environment is approximately 0-40 C. The robot is neither waterproof nor dustproof.

Avoid:

- rain or wet floors
- standing water
- gravel and dusty environments
- ice or very low-friction floors
- strong electromagnetic interference
- uncontrolled crowded spaces

Smooth tile and similar flooring can also reduce traction, particularly during aggressive motion.

## Battery

Published and internal documentation describes approximately two hours of general operating endurance, but actual duration depends on walking speed, pose, payload, terrain, and activity.

For storage, the lab manual recommends checking batteries periodically and, if charge falls below 30%, charging to approximately 70% before continued storage.

## Troubleshooting

### Robot jitters during SDK development

Check first for **controller conflict**:

- Is the built-in motion controller still active?
- Is the EDU actually in Develop mode?
- Are multiple DDS applications publishing actuator commands?

Then check:

- joint order
- gains
- command frequency
- units
- stale command data
- robot-model configuration

### Robot falls or cannot stand reliably

Check:

- floor friction
- battery state
- initialization state
- physical damage
- joint calibration status
- active control mode
- policy/controller compatibility

Do not recalibrate joints casually. The lab manual states that G1 is calibrated at the factory and recommends consulting Unitree support before recalibration for abnormal behavior.

### SDK cannot communicate

Check:

- selected network interface
- robot IP/subnet
- firewall
- DDS multicast/discovery
- SDK version
- robot firmware

### Perception appears incorrect

Check:

- camera/LiDAR obstruction
- coordinate transforms
- timestamps
- calibration
- sensor stream rates
- physical configuration of the robot

## Research Use

The G1 platforms are suitable for:

- humanoid locomotion
- reinforcement learning
- imitation learning
- learning from demonstration
- whole-body control
- manipulation
- dexterous manipulation on configured EDU units
- embodied AI
- mobile manipulation
- human-robot interaction
- manufacturing task research
- perception and sensor fusion

The EDU should be the default platform where experiments require SDK-level control or altered locomotion/manipulation policies. The Air is better treated as the less developer-focused configuration unless the specific purchased unit exposes the interface required by the experiment.

## References

- Unitree Robotics, *G1 User Manual V1.1*.
- Unitree Robotics, *G1 SDK Development Guide*.
- Unitree Robotics, *G1 Product Specifications*.
- `unitree_sdk2` developer documentation and examples.