Unitree G1-D Wheeled Humanoid

Unitree G1-D Wheeled Humanoid Robot

The G1-D emerged from a specific market need that the bipedal G1 cannot efficiently address: large-scale, sustained robotic data collection for training humanoid manipulation AI. Training the vision-language-action (VLA) models and imitation learning policies that enable robots to handle diverse real-world manipulation tasks requires millions of robot-object interaction episodes across diverse object categories, environments, and task variations. Generating this data through teleoperated bipedal robots is expensive and operationally constrained: bipedal platforms have limited operational endurance (1 to 2 hours per battery charge), require careful balance management during teleoperation, and impose operational overhead on the human teleoperation staff.

The G1-D addresses these constraints through its wheeled design. The Flagship's 6-hour battery life enables extended data collection sessions without frequent recharging interruptions. The wheeled base's inherent stability eliminates the balance overhead that bipedal teleoperation requires, reducing cognitive load on teleoperation staff and improving the quality of demonstration data. The wrist-mounted cameras capture the hand-object visual perspective that is most informative for training manipulation policies — a viewpoint that the head-mounted camera of many existing data collection systems cannot provide.

Unitree's timing with the G1-D also reflects the broader industry context of the "robot training data" market. As general-purpose humanoid AI becomes increasingly dependent on large datasets of human demonstrations, specialized data collection platforms that can efficiently generate high-quality robot demonstration data at scale represent a distinct and growing market segment from the general-purpose research humanoid segment.

The G1-D's integration into Unitree's data collection software framework — providing end-to-end tools for teleoperation data capture, annotation, training pipeline management, and policy deployment — positions it as a complete operational system rather than just a robot hardware platform.

Design and Physical Features

Wheeled Mobile Base: Standard vs. Flagship

The fundamental design distinction between the G1-D's two versions is the mobile base:

G1-D Standard: The upper body is mounted on a stationary base — no locomotion, suitable for fixed workstation data collection tasks where the robot is deployed at a specific location and handles objects brought to it. The stationary design simplifies deployment and reduces cost compared to the mobile Flagship.

G1-D Flagship: The upper body is mounted on a differential drive wheeled mobile base supporting movement at up to 1.5 meters per second. The base includes LiDAR sensors, depth cameras, and collision detection units — a full autonomous navigation sensor suite enabling the robot to navigate factory floors, logistics facilities, and laboratory environments without track guidance. The differential drive enables turning-in-place and arc navigation, covering the movement patterns needed for logistics, warehouse, and service data collection deployments.

Height Adjustability: 1260 to 1680 mm

The G1-D's adjustable height — between 1,260 and 1,680 mm — enables the upper body to match different workstation heights, counter heights, and shelving configurations without repositioning the entire mobile base. This height adjustability is critical for data collection across diverse task environments: the robot must be able to reach the specific height of the work surface where the task is performed to generate visually valid training data.

17 to 19 Degrees of Freedom

The G1-D's DOF count is lower than the bipedal G1 EDU family because it does not include the leg DOF that account for 12 of the bipedal G1's total. Without legs, the G1-D's upper body provides:

Standard (17 DOF): Arms and upper body without mobile base actuation DOF. Flagship (19 DOF): Arms and upper body with the mobile base DOF included in the total count.

The arm DOF configuration — with 7 DOF per arm confirmed by Gizmochina's launch coverage — provides the full shoulder-to-wrist workspace needed for manipulation tasks, and the wrist DOF enables the hand to approach objects from any orientation within the arm's reach envelope.

Binocular Head Camera and Dual Wrist Cameras

The G1-D's camera system is specifically designed for manipulation data collection quality:

HD binocular head camera: Provides stereo depth perception from the robot's egocentric viewpoint — the perspective from which an operator using VR teleoperation sees the workspace and the manipulated objects.

HD wrist cameras (one per arm): Mounted at the wrist, these cameras capture the near-field view of the hand-object interaction — the visual data most informative for training grasping and manipulation policies. Wrist cameras see the object from the same perspective the hand approaches it, providing the visual observation data that corresponds directly to the hand configuration and contact events that manipulation AI policies must learn from.

This three-camera configuration — binocular head plus bilateral wrists — captures the complete visual observation needed for both policy learning (what did the robot see when it made each decision?) and behavior cloning (what did the human demonstrator see when they performed each action?).

Technology and Specifications

G1-D Full Specifications

Modular End-Effector System

The G1-D's modular end-effector design is central to its data collection value proposition. Four end-effector configurations are supported:

Two-finger gripper: The simplest configuration for pick-and-place tasks requiring robust grasping of standardized objects. Suitable for logistics sorting, object transport, and basic manipulation data collection where full dexterity is not the research objective.

Dex3-1 three-finger hands (without tactile): Force-position controlled three-finger manipulation for data collection of grasping tasks across diverse object shapes, covering the grasping patterns that three-finger kinematic analysis enables.

Dex3-1 three-finger hands with tactile sensors: Same kinematic configuration with distributed contact pressure sensing — enabling data collection that captures not only visual and joint state observations but also the contact force distribution during grasps, providing richer training data for tactile-aware manipulation policy development.

Five-finger dexterous hands: For data collection of tasks requiring the full human grasp taxonomy — tool use, fine manipulation, and complex multi-finger coordination — where the three-finger configuration cannot reproduce the demonstrator's hand posture.

NVIDIA Jetson Orin NX (100 TOPS) and Data Collection Software

The Jetson Orin NX provides 100 TOPS of GPU-accelerated compute for on-device processing during data collection sessions: real-time visual perception, teleoperation policy assistance (where the robot's AI helps guide the teleoperator toward successful grasps), and data annotation (labeling captured frames with object categories, task stages, and success outcomes).

Unitree's data collection software framework — which integrates with the G1-D hardware — provides tools for managing the complete AI workflow: session management, synchronized multi-modal data recording (joint state, camera streams, tactile sensor data), data annotation workflows, and training pipeline integration. This full-stack software differentiates the G1-D from robots that capture raw data without an integrated pipeline for converting that data into trained AI models.

Roller Skate and Ice Skate Demonstrations: Adaptive Locomotion Interface

The April 2026 demonstration of the G1-D on roller skates and ice skates reveals a design characteristic of the G1-D's wheeled-leg integration: the control architecture is generalized enough to manage balance and locomotion across different wheel-like locomotion interfaces. On roller skates, the robot executes 360-degree turns and one-leg spins; on ice skates, it maintains balance on narrow blade contacts. A front flip — demonstrated on wheels — extends the athletic capability into the vertical plane.

Interesting Engineering's coverage quotes Unitree: "Humanoid robots are the ideal form of general-purpose robots (perfect for general AI and human-derived data). They can work without wheels — but they can also have wheels if they want. Whatever works." This design philosophy — agnosticism between wheeled and legged locomotion — reflects the G1-D's positioning as a platform where capability coverage matters more than locomotion purity.

Applications and Use Cases

Large-Scale Manipulation Data Collection

The G1-D's primary designed application is generating manipulation training data for humanoid AI. Research institutions and AI companies developing general-purpose robot manipulation policies require millions of labeled robot demonstration episodes across hundreds of object categories and task types. The G1-D's 6-hour Flagship battery, stable wheeled base, multi-camera configuration, and integrated data management software enable sustained data collection workflows at a scale that bipedal teleoperation systems cannot efficiently achieve.

Industrial and Service Task Execution

The G1-D Flagship's 1.5 m/s mobile base, LiDAR navigation, and modular end-effectors make it suitable for deployment in industrial logistics, warehouse operations, and service environments where the robot must navigate between workstations and perform manipulation tasks across a facility. Applications include inventory inspection, object sorting and transport, shelf restocking, and quality inspection support.

Warehouse Automation

The Flagship's combination of autonomous navigation (LiDAR + depth camera mobile base) and modular manipulation end-effectors addresses warehouse automation tasks — order picking, bin handling, parcel sorting — in a wheeled humanoid form factor that navigates standard warehouse aisles without the terrain limitations of legged-only platforms.

Retail and Consumer Interaction

With the five-finger dexterous hand option and HD visual system, the G1-D is suitable for retail environments where the robot interacts with customers, retrieves products from shelves, and handles the full diversity of retail product shapes and packaging materials. The wheeled base's stability compared to bipedal operation provides reliability for sustained retail deployment.

Research: Wheeled-Legged Hybrid Locomotion Control

The roller skate and ice skate demonstrations position the G1-D as a research substrate for wheeled-legged hybrid locomotion control — an increasingly active research area studying how robots can efficiently coordinate wheel and leg actuation for different terrain and surface conditions. For research teams studying this hybrid locomotion domain, the G1-D provides a commercially available platform with demonstrated performance in unconventional locomotion interfaces.

Advantages and Benefits

6-Hour Flagship Battery for Full-Shift Data Collection: The Flagship's 6-hour runtime enables complete shift-length data collection sessions without mid-session battery interruptions — more than triple the G1 bipedal's 2-hour runtime for the same platform size.

Stable Wheeled Base Eliminates Bipedal Balance Overhead: Wheeled locomotion eliminates the computational and operational overhead of continuous bipedal balance management, enabling the robot's full compute capacity and the teleoperator's full attention to be devoted to manipulation quality rather than balance management.

Dual Wrist Cameras for High-Quality Manipulation Training Data: The wrist-mounted cameras capture the hand-object visual perspective that directly corresponds to manipulation policy inputs — a critical data quality advantage for training generalizable grasping AI.

Height-Adjustable Upper Body for Multi-Workstation Deployment: The 1,260 to 1,680 mm height range enables the G1-D to match diverse workstation heights across different data collection sites without hardware modification.

Modular End-Effectors for Task-Specific Data Collection: The four end-effector options — two-finger, Dex3-1, Dex3-1+tactile, five-finger — configure the platform for data collection across different manipulation research objectives within the same hardware investment.

Full-Stack Data Collection Software Framework: The integrated software for session management, multi-modal data recording, annotation, and training pipeline integration reduces the operational overhead of running large-scale data collection programs.

Comparison: G1-D vs. Bipedal G1 EDU

The G1-D's longer battery life, wrist cameras, and data-collection-optimized software make it the superior platform for AI training data generation. The bipedal G1 EDU's stairs and terrain capability make it superior for research requiring outdoor locomotion and multi-level environment navigation.

Frequently Asked Questions (FAQ)

What is the Unitree G1-D? The Unitree G1-D is Unitree Robotics' first wheeled humanoid robot, launched in November 2025. It combines a human-like upper body with dual 7-DOF arms on a wheeled mobile base, designed primarily for large-scale data collection, AI model training, and real-world task execution. It is available in Standard (stationary, 17 DOF) and Flagship (mobile differential drive at 1.5 m/s, 19 DOF, up to 6-hour battery) versions. Height adjusts between 1,260 and 1,680 mm. It includes HD binocular head cameras and HD wrist cameras, NVIDIA Jetson Orin NX (100 TOPS), and supports modular end-effectors including two-finger grippers, Dex3-1 three-finger hands with or without tactile sensors, and five-finger dexterous hands.

How does the Unitree G1-D differ from the bipedal G1? The bipedal G1 uses legs for locomotion, providing stair climbing, outdoor terrain navigation, and up to 43 DOF across the full body at approximately 2 hours of battery runtime. The G1-D uses a wheeled base for locomotion, providing up to 1.5 m/s on flat surfaces with up to 6 hours of battery runtime at 17 to 19 DOF in the arms and upper body. The G1-D adds wrist cameras for manipulation data quality that the bipedal G1 lacks, and integrates with a dedicated data collection and AI training software framework. The G1-D is optimized for sustained data collection and task execution on flat industrial floors; the bipedal G1 is optimized for research requiring terrain versatility.

What did the Unitree G1-D roller skate and ice skate demonstration show? An April 2026 video showed the G1-D performing on roller skates and ice skates — executing 360-degree turns, one-leg spins, and front flips on wheels, and maintaining balance on narrow ice skate blade contacts. The demonstration illustrates the G1-D's wheeled-legged control architecture's ability to generalize balance and locomotion coordination to different wheel-like locomotion interfaces beyond its standard wheeled base, and reflects Unitree's design philosophy that humanoid robots should be able to use "whatever works" for locomotion in different contexts.

What end-effectors does the G1-D support? The G1-D supports four modular end-effector configurations: a two-finger gripper for robust pick-and-place of standardized objects; Dex3-1 three-finger dexterous hands without tactile sensors; Dex3-1 three-finger hands with integrated tactile sensor arrays; and five-finger dexterous hands for full human grasp taxonomy coverage. The modular design allows the same G1-D base platform to be configured for different data collection research objectives by swapping end-effectors.

Summary

The Unitree G1-D Wheeled Humanoid Robot represents Unitree's first purpose-designed data collection and task execution humanoid platform — combining a human-like upper body with 7-DOF dual arms, HD binocular head cameras, bilateral HD wrist cameras, NVIDIA Jetson Orin NX (100 TOPS) compute, and a modular end-effector system on a wheeled mobile base that achieves up to 1.5 m/s and up to 6 hours of operational endurance in the Flagship configuration. Launched in November 2025 and demonstrated performing on roller skates and ice skates in April 2026, the G1-D addresses the large-scale AI training data generation challenge that bipedal humanoid platforms cannot efficiently solve due to their limited battery endurance and balance management overhead. For AI research organizations, manufacturing firms developing humanoid robot policies, and data collection service providers needing a sustained, stable, wrist-camera-equipped manipulation data platform, the G1-D represents a commercially available and well-supported entry into the specialized wheeled humanoid data platform category.