Menu

UNITREE G1-D Flagship Humanoid Robot

UNITREE G1-D Flagship Humanoid Robot
UNITREE G1-D Flagship Humanoid Robot
UNITREE G1-D Flagship Humanoid Robot
New

The Unitree G1-D Flagship is a wheeled humanoid service robot engineered for sustained commercial deployment, combining dual 7 DoF arms, an adjustable-height telescoping column spanning 1260 to 1680 mm, and an NVIDIA Jetson Orin NX processor delivering 100 TOPS of on-board AI inference. Weighing approximately 80 kg with its dual-battery architecture, the platform sustains up to 6 hours of chassis autonomy and achieves a maximum operational reach of ~2 m.

Total DoF (excl. End Effector) 19 (7×2 arms + 2 waist + 1 column + 2 base)
Adjustable Height Range 1260–1680 mm (max working height ~2 m)
AI Computing Module NVIDIA Jetson Orin NX — 100 TOPS
Chassis Battery Life ~6 h (30 Ah integrated battery)

Unlike fully bipedal humanoids that spend engineering budget on dynamic balance, the G1-D Flagship redirects that effort entirely into arm capability and task intelligence. The wheeled differential-drive chassis removes bipedal stabilisation complexity, enabling the upper body to focus on dexterous manipulation and perception. The image below captures the G1-D operating as an autonomous barista — managing espresso equipment with calm precision, a scenario that demands both careful object handling and reliable positional repeatability.

Unitree G1-D Flagship humanoid robot working autonomously as a barista in a coffee shop, standing behind a counter next to a professional Rocket espresso machine

Telescoping Column: From Floor Level to a 2 m Working Envelope

The most consequential mechanical choice in the G1-D's design is its telescoping lifting column. The column travels 450 mm of vertical stroke at up to 60 mm/s, controlled to 1 mm positioning accuracy. At minimum height the robot stands 1260 mm tall — compact enough to handle floor-level tasks and fit through standard doorways. At maximum extension it reaches 1680 mm, pushing the arm workspace up to ~2 m above the ground. That range covers virtually every shelf height found in retail, warehouse, and laboratory environments. The waist simultaneously articulates across Z±155° and Y -2.5° to +135°, allowing the arms to sweep from below the chassis to well above head height without repositioning the chassis.

The diagram below illustrates both the vertical stroke numbers and the full waist range of motion that together define the G1-D's expanded operational workspace — showing the robot handling a standard shipping carton at conveyor height with a single-arm grasp.

Unitree G1-D Flagship expanded operational workspace diagram showing vertical workspace 0–2 m, waist ROM Z ±155°, waist ROM Y -2.5° to +135°, robot lifting a cardboard box on a warehouse conveyor belt

7 DoF Arms and a Modular End Effector Ecosystem

Each arm carries 7 active degrees of freedom: shoulder pitch, shoulder roll, shoulder yaw, elbow, wrist roll, wrist pitch, and wrist yaw. Seven-DoF arms are the gold standard in professional manipulation research precisely because they allow continuous null-space motion — the robot can reorient the wrist without moving the end effector, which matters enormously in confined workspaces. With a single-arm payload of ~3 kg and a reach of ~0.45 m, the arms handle object weights encountered in retail, food service, logistics, and light assembly.

The end effector is intentionally not fixed. Four hardware options are available depending on the application: a 2-finger force-controlled gripper for general object handling, a 3-finger dexterous hand without tactile sensing, a 3-finger dexterous hand with tactile sensing for tasks requiring contact feedback, and a 5-finger dexterous hand for the most human-like manipulation requirements. The annotated diagram below shows both height configurations and the full range of compatible end effectors in context.

Annotated diagram of Unitree G1-D Flagship showing two height configurations, head-mounted HD binocular camera, wrist-mounted HD cameras, optional end effectors (2-finger gripper, 3-finger and 5-finger dexterous hands), adjustable height mechanism 1260–1680 mm, and optional mobile chassis with max speed 1.5 m/s

19 DoF Platform: The Kinematic Architecture Explained

The body's 19 total degrees of freedom (excluding end effectors) are distributed with deliberate precision. The arms account for 14 DoF (7 per arm), giving each limb the same kinematic redundancy found in professional collaborative robot arms. Two waist DoF — rotation about the Z-axis and pitch about the Y-axis — allow the torso to twist and bend independently, decoupling arm positioning from chassis heading. One column DoF handles vertical height adjustment. Two base DoF correspond to the chassis differential drive: forward/backward velocity and yaw angular velocity. When a 2-finger gripper is added to each arm, the total rises to 21 DoF.

The platform specification diagram below summarises the complete DoF breakdown per subsystem, confirming that the arm architecture alone — at 7 DoF per limb — matches the kinematic capability of a standalone industrial collaborative arm.

Higher-DOF Robot Platform infographic for Unitree G1-D showing total DoF 19, arm DoF 7×2, waist DoF 2, column DoF 1, base DoF 2

Lower-Latency Control: VR Teleoperation and Precision Positioning

High-fidelity teleoperation is the primary mechanism for collecting demonstration data that trains autonomous policies. The G1-D's control pipeline delivers a system teleoperation latency of <100 ms with a 60 Hz sampling rate — fast enough for a human operator wearing a VR headset to maintain a convincing sense of embodiment during dexterous manipulation. Column positioning accuracy under VR teleoperation reaches ±0.5 mm; end-effector gripper accuracy is ±0.1 mm (accuracy varies with end-effector configuration). These tolerances matter for tasks like picking small components, inserting connectors, or folding flexible materials where imprecise reproduction would invalidate the training demonstration.

The image below shows an operator performing VR-guided teleoperation alongside the G1-D, illustrating the system's key control response parameters that make high-quality data collection practical at scale.

VR teleoperation session with Unitree G1-D Flagship — operator wearing VR headset controlling the robot, showing system teleoperation latency below 100 ms, sampling rate 60 Hz, lifting accuracy ±0.5 mm, gripper accuracy ±0.1 mm

Wheeled Chassis and Autonomous SLAM Navigation

The mobile base operates on a differential-drive with two independent drive wheels, supporting 360° in-place rotation and a maximum travel speed of 1.5 m/s. Onboard chassis sensors include a 3D LiDAR, two depth cameras, two physical collision sensors, and two low-obstacle detection sensors — a sensing suite that enables the robot to build maps, localise itself, avoid dynamic obstacles, and autonomously return to its charging station. Navigation is managed through a SLAM service accessible via REST API, with support for points of interest, virtual walls, forbidden zones, and multi-point route sequencing. The robot's integrated 30 Ah battery powers the chassis for approximately 6 hours before docking is required. The image below shows the G1-D performing a folding task in a domestic bedroom — an illustrative scenario where the chassis navigates to the workspace and the column adjusts to precisely match the surface height.

Unitree G1-D Flagship on wheeled mobile chassis performing a clothes-folding task in a domestic bedroom, demonstrating precision arm manipulation at bed height
Expert Verdict: The G1-D Flagship occupies a distinct category from both bipedal humanoids and fixed collaborative arms. Its telescoping column solves the height-coverage problem that trips up single-height mobile manipulators, while the 7 DoF arm architecture closes the dexterity gap compared to earlier service robot platforms. The combination of <100 ms teleoperation latency, a 6-hour chassis battery, and a full SLAM navigation stack means this platform can meaningfully run production-grade demonstration data collection — not just laboratory trials. One practical note for deployers: when configuring the SLAM map, mark all charging dock and POI positions during the initial mapping pass at the lowest column height; remapping with a different column height shifts the sensor elevation and can introduce localisation drift in environments with low feature density.

End-to-End AI Platform: From Data Acquisition to Deployed Policy

The G1-D is not simply a robot — it is the physical execution node of a complete embodied AI development stack. Unitree's platform integrates three interconnected layers: a streamlined data acquisition pipeline, a comprehensive model training and inference environment, and the UnifoLM-WMA-0 world-model–action architecture. The data pipeline standardises collection across multiple robot platforms using visual template management, one-click task generation, high-concurrency scheduling for hundreds of simultaneous robots, and 24/7 continuous collection — all feeding directly into mainstream training formats.

The image below shows the G1-D operating on an industrial assembly line alongside multiple identical units — a high-throughput data collection scenario enabled by the platform's concurrent architecture.

Multiple Unitree G1-D Flagship robots operating on an industrial conveyor belt as part of an end-to-end data collection and training platform for humanoid AI

UnifoLM-WMA-0: World-Model–Action Architecture

The model training layer supports distributed training with up to 90% GPU utilisation, integration with open-source models including PI and GROOT, one-click model deployment, and a high-fidelity simulation environment for policy evaluation before physical rollout. At the core is UnifoLM-WMA-0 — Unitree's open-source world-model–action architecture spanning multiple robot embodiments. It operates in two modes: a decision-making mode that predicts future physical interactions to guide policy execution, and a simulation mode that generates high-fidelity synthetic training data from robot motion inputs. The full Sim2Real pipeline is documented and supported, covering architecture selection, training configuration, real-time monitoring, parameter editing, simulation testing, and model deployment.

Service, Retail, and Industrial Inspection Use Cases

The G1-D's variable reach and SLAM autonomy make it practically deployable across three main verticals. In retail environments, the robot can navigate store aisles, identify shelf positions using onboard cameras, and restock products without human assistance. The image below shows the G1-D handling packaged goods on a bulk food shelf — a task that demands accurate object identification, controlled force application, and reliable spatial positioning.

Unitree G1-D Flagship stocking shelves in a retail bulk food section, using its dexterous arms to handle packaged products on high shelves

In industrial and data-centre contexts, the G1-D's ability to navigate crowded aisles, extend its column to reach high rack positions, and apply controlled force through its arms makes it a viable tool for equipment inspection, cable management, and component handling. The image below shows the platform working in a server room — a space characterised by narrow aisles, vertically stacked hardware, and the need for very precise, non-contact-safe arm movements.

Unitree G1-D Flagship performing industrial inspection and maintenance on server rack equipment in a data centre, using its adjustable-height column and 7 DoF arm

Technical Specifications of the Unitree G1-D Flagship

Mechanical Dimensions

Model G1-D Flagship
Overall Dimensions (Min. Column Height) ~1260 × 525 × 570 mm
Overall Dimensions (Max. Column Height) ~1680 × 525 × 570 mm
Total Weight (incl. battery) ~80 kg
Cooling System Local air cooling

Degrees of Freedom

Total DoF (excl. End Effector) 19
Single Arm DoF (excl. End Effector) 7
Waist DoF 2
Column DoF 1
Base DoF 2
Total DoF with 2-Finger Gripper ×2 21 (19 + 1 per gripper × 2)

Arm Performance

Max. Single Arm Payload ~3 kg
Arm Reach (excl. End Effector) ~0.45 m
End Effector Options 2-Finger Gripper / 3-Finger Dexterous Hand (No Tactile) / 3-Finger Dexterous Hand (With Tactile) / 5-Finger Dexterous Hand

Column & Waist Range of Motion

Column Lifting Travel 450 mm
Column Lifting Speed Max. 60 mm/s
Column Lifting Accuracy (general) 1 mm
Lifting Accuracy (VR teleoperation) ±0.5 mm
Max. Working Height ~2 m
Waist Joint Range — Z-axis ±155°
Waist Joint Range — Y-axis -2.5° to +135°

Chassis Performance

Chassis Dimensions (L × W × H) 570 × 525 × 197 mm
Drive Type Differential drive — supports 360° in-place rotation
Maximum Mobility Speed 1.5 m/s
Chassis Sensors LiDAR ×1 + Depth Camera ×2 + Physical Collision Sensor ×2 + Low-Obstacle Detection Sensor ×2

Computing & AI

Basic Computing Power 8-core high-performance CPU
High-Performance Computing Module NVIDIA Jetson Orin NX 16 GB (100 TOPS)

Sensors & Perception

Head HD Binocular Camera ×1 — FOV: H 115°, V 80°, D 125° — Resolution: 3840 × 1200
Wrist HD Camera ×2 — FOV: H 130°, V 60°, D 160° — Resolution: 1920 × 1080
Base LiDAR ×1
Base Depth Camera ×2
Physical Collision Sensor (Base) Present
Low-Obstacle Detection Sensor (Base) Present

Audio & Interaction

Microphone Array 4-mic linear array, 20 mm spacing
Speaker 8 Ω 3 W (5 W peak)
RGB Light Strip 256 colors
ASR (Speech Recognition) Local offline model
TTS (Text-to-Speech) Local offline synthesis — Chinese & English

Connectivity

WiFi WiFi 6
Bluetooth Bluetooth 5.2

Control & Teleoperation

System Teleoperation Latency <100 ms
Sampling Rate 60 Hz
End-Effector Gripper Accuracy ±0.1 mm (varies with end-effector configuration)

Power & Battery — Upper Body

Upper Body Power Supply Battery or direct cable connection
Upper Body Battery Capacity (quick-release) 9000 mAh
Upper Body Battery Life ~2 h
Upper Body Charger 54 V / 5 A

Power & Battery — Chassis

Chassis Power Supply Battery / charging station
Chassis Battery Capacity (integrated) 30 Ah
Chassis Battery Life ~6 h
Chassis Charging Station 51 V / 10 A

Software & Development

SDK unitree_sdk2 (based on G1 SDK)
SLAM Navigation Yes — REST API, supports POIs, virtual walls, forbidden zones, multi-point routes
AI World Model UnifoLM-WMA-0 (open-source)
Supported Open-Source Models PI, GROOT and community datasets
Distributed GPU Training Utilisation Up to 90%
Video Streaming ZMQ / WebRTC (head b
Robot Specifications
Navigation & Sensors LiDAR Depth Camera: 2 Physical Collision Sensor: 2 Low-Obstacle Detection Sensor: 2
Robot Type Humanoid
Application / Purpose Service/Hotel/Restaurant/Cafè
Max Payload (kg) 3
Max Travel Speed (m/s) 1.5
Battery Life (h) 6
Computing Module NVIDIA Jetson Orin NX 16GB(100TOPS)
SDK / Secondary Development Yes
Weight and Dimensions
Gross Weight (kg) 80

Write a comment