Menu

UNITREE G1-COMP EDU / COMPETITION Humanoid Robot

UNITREE G1-COMP EDU / COMPETITION Humanoid Robot
UNITREE G1-COMP EDU / COMPETITION Humanoid Robot
New

The Unitree G1-Comp EDU is a full-size bipedal humanoid robot engineered specifically for competitive robotics and advanced university research, standing approximately 130 cm tall and weighing around 35 kg. With 37 degrees of freedom, a maximum joint torque of 120 N.m, integrated Dex3-1 force-controlled dexterous hands, and a dedicated NVIDIA Jetson Orin NX development module delivering 100 TOPS of AI inference, the G1-Comp EDU sets a new benchmark for deployable humanoid platforms.

Total Degrees of Freedom 37 (G1-Comp EDU configuration)
Maximum Knee Torque 120 N.m
AI Computing Module NVIDIA Jetson Orin NX — 100 TOPS
Battery Life ~2 h (9000 mAh, quick-release)

The image below captures the G1-Comp in its natural competitive environment — a full-size football pitch — demonstrating the stable gait and purposeful motion that earned it the label "Football Icon Designed for Competitions."

Unitree G1-Comp EDU humanoid robot on outdoor football stadium field, about to kick a ball

Competition-Built Body: 130 cm, 35 kg, Ready to Move

At 1320 × 450 × 200 mm (standing) and foldable to 690 × 450 × 300 mm for transport, the G1-Comp EDU occupies the physical envelope of a small adult human. Weighing approximately 35 kg with battery, the frame achieves a strong power-to-weight ratio thanks to an all-electric drivetrain. No hydraulics, no pneumatics — every joint is driven by Unitree's proprietary hollow-shaft motor, and all wiring runs internally through the hollow joint structure, eliminating external cable snag points entirely.

The dual-view below confirms the standing and folded dimensions that make the G1-Comp straightforward to transport between competition venues and research laboratories.

Unitree G1-Comp EDU full body standing pose with body size labels: weight ~35 kg, height ~130 cm

37 Degrees of Freedom: Human-Like Movement Architecture

Humanoid mobility is only as credible as its kinematic architecture. The G1-Comp EDU ships with 37 total degrees of freedom — the base platform contributes 25 to 45 configurable DoF across legs, waist, arms, and head, while the bundled Dex3-1 three-finger dexterous hands add 7 DoF per hand plus 2 optional wrist DoF per arm. Each leg articulates across 6 DoF (Hip 3 + Knee 1 + Ankle 2), giving the robot the hip roll, yaw, and pitch range needed for agile bipedal walking, side-stepping, and recovery from external disturbances.

Extra-Large Joint Range of Motion

Compared to many academic humanoid platforms, the G1-Comp offers an unusually wide angular envelope. The waist rolls on Z±155°, X±45°, Y±30°; the knee flexes from 0° to 165°; hip pitch spans ±154°. These ranges allow crouching, reaching, and dynamic weight-shifting manoeuvres that constrained-DoF robots simply cannot execute. The shell, visible below, is constructed from aluminium alloy reinforced with high-strength engineering plastics — a combination that absorbs impact loads in competition without contributing unnecessary mass to the kinematic chain.

Close-up of Unitree G1-Comp robot torso and arms showing aluminium alloy and high-strength engineering plastic shell construction

Dual-Encoder Precision on Every Joint

Position accuracy under load is a persistent challenge for high-DoF humanoids. The G1-Comp addresses this with a dual-encoder system on every joint — one encoder on the motor rotor, a second on the output shaft. This redundant feedback loop delivers accurate and stable joint states even when external interference or mechanical backlash would degrade single-encoder performance. The result: reliable Sim2Real policy transfer, where behaviour trained in Isaac Gym or MuJoCo executes predictably on the physical robot.

Dual encoder system diagram for Unitree G1-Comp — accurate and stable joint control, no fear of interference

Locomotion Performance: 2 m/s Bipedal Gait

The G1-Comp's onboard motion control stack delivers a peak locomotion speed of 2 m/s — fast enough to compete in standardised RoboCup soccer match formats. The industry-leading motion controller is purpose-built for competitive match environments, maintaining gait stability under ball-contact perturbations and lateral crowding from opposing robots.

Super-Stable Balance Control

In practice, balance control determines whether a match robot stays upright or ends up on the turf. The G1-Comp's control system absorbs unexpected pushes and uneven surface variations without breaking gait continuity. The frame below shows the robot mid-stride with its centre of mass correctly projected over the support polygon — the hallmark of a well-tuned whole-body controller.

Unitree G1-Comp humanoid robot demonstrating super-stable balance and gait control on indoor artificial turf

Omnidirectional Walk

Beyond straight-line locomotion, the G1-Comp supports omnidirectional walking — translating laterally, rotating in place, and changing heading without stopping. This capability is essential for goal-side repositioning in football and equally useful for inspection or manipulation tasks in research settings where obstacle-rich environments demand agile footprint management.

Unitree G1-Comp EDU demonstrating omnidirectional walk capability in front of football goal posts on indoor pitch

Dex3-1 Dexterous Hands: Force-Controlled Manipulation

Unlike many competition humanoids that ship with passive grippers, the G1-Comp EDU includes a pair of Dex3-1 three-finger dexterous hands as standard equipment. Each hand provides 7 active degrees of freedom: the thumb contributes 3 DoF; the index finger and middle finger each contribute 2 DoF. The force-sensing array spans a perception range of 10 g to 2500 g, enabling precise grasping of both delicate objects and standard competition equipment. Operating voltage is 12–58 V, drawing directly from the robot's power bus. An optional tactile sensor array upgrade is available for research workflows requiring skin-level contact feedback.

The portrait below shows the complete G1-Comp EDU system — note the distinctive helmet-mounted depth camera, the bimanual arm configuration, and the Dex3-1 gloved-appearance hands that give the platform its distinctly humanoid character.

Unitree G1-Comp EDU humanoid robot full upper body portrait showing helmet-mounted depth camera and dexterous hand configuration
Expert Verdict: The G1-Comp EDU occupies a rare position: it is simultaneously a competition-certified RoboCup-class platform and a fully open research tool. The combination of NVIDIA Jetson Orin NX, dual-encoder joints, YOLO11 visual recognition, and Dex3-1 dexterous hands in a single 35 kg package is compelling. Researchers who previously had to choose between mechanical capability and software openness no longer face that trade-off. One practical note: when attaching the Dex3-1 hands for manipulation tasks, add a modest outward shoulder-motor offset (3–5°) to prevent interference between the dexterous hand thumb and the robot's lateral torso surface — a tip documented in Unitree's own SDK release notes.

NVIDIA Jetson Orin NX: 100 TOPS for On-Board AI

The dedicated development computing unit — an NVIDIA Jetson Orin NX — provides 100 TOPS of AI inference performance alongside an 8-core Arm Cortex-A78AE CPU clocked at up to 2 GHz, 16 GB of unified memory, and 1024 NVIDIA Ampere GPU cores. This is not a thin client offloading compute to the cloud; all inference, perception, and control decisions run on-board in real time. The operational computing unit handling low-level motor control runs a separate Unitree-proprietary stack that is not accessible to end users, preserving motion control integrity while leaving the full Jetson environment open for custom development.

Hardware Interface Configuration

The G1-Comp's right-side panel exposes a rich hardware interface for secondary development. The image below shows the full configuration: USB Type-C ports supporting USB 3.0 and USB 3.2 host modes with 5 V / 1.5 A power output, dual Gigabit Ethernet RJ45 ports for high-bandwidth sensor feeds, and multi-voltage power rails at 5 V, 12 V, 24 V, and 54.8 V. The depth camera system uses an Intel RealSense D455, combined with 2-DoF head rotation to achieve 180° field of view coverage. The 4-microphone array incorporates noise reduction and echo cancellation for reliable voice command reception even in noisy competition environments.

Unitree G1-Comp hardware configuration panel showing USB, Gigabit Ethernet, vision (Intel RealSense D455), audio array, and gripper compatibility

Open Development Ecosystem

Beyond raw compute, the G1-Comp EDU ships with a comprehensive software stack. The diagram below maps all six pillars of the development ecosystem: a multi-level API layer (high-level, low-level, DDS, audio/lighting); simulation environments in Isaac Gym and MuJoCo; multimodal interaction via the UnifoLM large language model with TTS and ASR support; ROS ecosystem compatibility; a mobile APP for rapid configuration; and the Jetson Orin NX as the development computing unit. ROS compatibility means existing lab code bases, sensor drivers, and visualisation pipelines port with minimal rework.

Unitree G1-Comp software ecosystem diagram: API interfaces, Isaac Gym / MuJoCo simulation, UnifoLM multimodal interaction, ROS support, APP control, and Jetson Orin NX development computing unit

RoboCup SDK: Sim2Real from Training to Competition

The dedicated RoboCup SDK bridges the gap between policy training and match-day deployment. Three specialised API layers address every stage of a competition agent's pipeline. The Visual Recognition API exposes the built-in YOLO11 real-time object detection network, providing a rich stadium information interface that identifies ball position, goal orientation, teammate and opponent locations. The Spatial Positioning API combines monocular geometric positioning with binocular depth positioning for accurate metric pose estimates within the pitch. The Motion Control API translates high-level decision signals — derived from visual and positional inputs — into valid locomotion and manipulation commands.

Training itself is supported by the unitree_rl_gym reinforcement learning framework, which integrates Isaac Gym and MuJoCo for efficient physical simulation. Training parameters (number of parallel environments, random seed, maximum iterations) are fully configurable. The complete Sim2Sim → Sim2Real pipeline is documented and supported, enabling teams to iterate rapidly in simulation before deploying policies on the physical robot. The image below shows the G1-Comp executing a trained football-approach behaviour on an indoor test pitch — an entirely hardware-in-the-loop validation scenario.

Unitree G1-Comp EDU humanoid robot approaching a football on an indoor test pitch during reinforcement learning policy validation
Tech Tip: When transitioning a trained policy from Isaac Gym simulation to the physical G1-Comp, enable the sim-to-real domain randomisation presets for terrain friction (μ = 0.4–1.2) and motor delay (5–20 ms). These parameter ranges reflect the actual variability observed on artificial turf competition surfaces. Training without domain randomisation produces policies that degrade significantly on the first physical roll-out.

9000 mAh Smart Battery: ~2 Hours of Continuous Operation

The G1-Comp runs on a 9000 mAh, 13-string lithium battery that delivers approximately 2 hours of operational autonomy under typical mixed-activity conditions. Quick-release mechanics mean battery swaps take seconds rather than minutes — critical in tournament settings where turnaround time between matches is limited. The charger operates at 54 V / 5 A. Intelligent OTA firmware updates deploy over-the-air, keeping motion control and SDK components current without requiring physical connection to a laptop.

The graphic below illustrates the approximately 120-minute runtime figure along with the quick-change and fast-charging features included in the G1-Comp package.

Unitree G1-Comp EDU battery life infographic showing approximately 120 minutes runtime with quick change and fast charging support

Technical Specifications of the Unitree G1-Comp EDU

The annotated diagram below provides a visual reference for the key hardware components and their physical placement on the G1-Comp EDU platform before the full numerical specification tables.

Annotated technical specification diagram of the Unitree G1-Comp EDU showing component labels for head DoF, depth camera, microphone array, speaker, arm DoF, core sports module, hollow joint wiring, athletic ability, single leg freedom, and quick-release battery

Mechanical Dimensions

Model G1 Comp
Height × Width × Thickness (standing) 1320 × 450 × 200 mm
Height × Width × Thickness (folded) 690 × 450 × 300 mm
Weight (with battery) ~35 kg+
Shell Material Aluminium alloy + high-strength engineering plastics
Calf + Thigh Length 0.6 m
Arm Span ~0.45 m

Degrees of Freedom

Total DoF — G1-Comp EDU configuration 37
Total DoF — base platform range 25–45 (configurable)
Single Leg DoF 6 (Hip 3 + Knee 1 + Ankle 2)
Waist DoF 1 + optional 2 additional
Single Arm DoF 5 (Shoulder 3 + Elbow 2)
Head DoF 2
Single Hand DoF — Dex3-1 (included) 7 + optional 2 wrist DoF (Thumb 3 + Index 2 + Middle 2)

Joint Performance

Maximum Torque — Knee Joint 120 N.m
Arm Maximum Load ~3 kg
Maximum Locomotion Speed 2 m/s
Joint Encoder Type Dual encoder (rotor + output shaft)
Full Joint Hollow Electrical Routing Yes — no external cables
Cooling System Local air cooling

Joint Movement Range

Waist Joint Range Z±155°, X±45°, Y±30°
Knee Joint Range 0~165°
Hip Joint Range P±154°, R-30~+170°, Y±158°
Wrist Joint Range P±92.5°, Y±92.5°
Head Joint Range P:-90°~+22.7°, Y:-50°~+50°

Computing & AI

Basic Computing Power 8-core high-performance CPU
Development Computing Module NVIDIA Jetson Orin NX
AI Performance 100 TOPS
Jetson CPU Arm Cortex-A78AE, 8 cores, up to 2 GHz
Jetson GPU 1024 NVIDIA Ampere architecture CUDA cores
Jetson Memory 16 GB unified memory

Sensors & Perception

Depth Camera Intel RealSense D455 (180° FOV with head rotation)
3D LiDAR Yes — 360° horizontal FOV, 59° vertical
Microphone Array 4-mic array with noise reduction and echo cancellation
Speaker 5 W stereo

Connectivity & Interfaces

WiFi WiFi 6
Bluetooth Bluetooth 5.2
Wired Network Gigabit Ethernet ×2 (RJ45)
USB Interfaces USB 3.0 Type-C ×3, USB 3.2 / DP1.4 Type-C ×1
Power Rails (developer-accessible) 5 V, 12 V, 24 V, 54.8 V

Power & Battery

Power Supply 13-string lithium battery
Smart Battery Capacity 9000 mAh (quick-release)
Charger 54 V / 5 A
Battery Life ~2 h

Dex3-1 Three-Finger Dexterous Hand

Total DoF per Hand 7 active (Thumb 3 + Index 2 + Middle 2)
Force Sensing Range 10–2500 g
Operating Voltage 12–58 V
Thumb Joint Angles 0°~+100°, -35°~+60°, -60°~+60°
Index and Middle Finger Angles 0°~+90°, 0°~+100°
Tactile Sensor Arrays Optional — 9 array sensors per hand

Software & Development

Simulation Environments Isaac Gym, MuJoCo
Reinforcement Learning Framework unitree_rl_gym (Sim2Sim + Sim2Real)
Visual Recognition
Robot Specifications
Ingress Protection (IP) LiDAR 3D, Depth Camera
Robot Type Humanoid
Application / Purpose Education, R&D Platform
Max Payload (kg) 3
Max Travel Speed (m/s) 2
Battery Life (h) 2
SDK / Secondary Development Yes

Write a comment