Unitree's wheeled humanoid robot has quietly launched, marking another step forward in its operational capabilities.

Today, Unitree Technology added a full-stack solution for data collection and training of humanoid robots to its official website. This solution consists of a high-performance humanoid robot body, systematic data collection tools, and comprehensive model training and inference tools. The body refers to its new wheeled humanoid robot, the G1-D.

G1-D is Unitree's third general-purpose humanoid application solution, suitable for service, lifestyle, commercial, and industrial scenarios. The previous two were the Unitree G1-Comp and the Iron Fist King, respectively applied to soccer matches and boxing events.

Compared to bipedal forms, wheeled robots are characterized by their stability and reliability. Their advantages, such as not easily falling in flat environments, reduced energy consumption, and adjustable height, allow for rapid deployment in scenes like workshops. This is also an important reason why the industry currently adopts more wheeled robots in factory settings.

Hardware Configuration Balances Flexibility and Practicality

As Unitree's first wheeled humanoid robot, the G1-D is available in a standard version and a flagship version. The standard version does not have a mobile chassis. The flagship version's mobile chassis has a speed of 1.5 m/s, can rotate 360° in place, and offers about 6 hours of battery life. The standard version has a battery life of approximately 2 hours.

G1-D Flagship Edition features a total of 19 degrees of freedom (excluding end-effectors), with each arm having 7 degrees of freedom. The maximum payload per arm is 3 kg, accommodating various operational requirements. Both the waist and chassis have 2 degrees of freedom, while the pillar has 1 degree of freedom.

In terms of workspace, G1-D adopts a mobile lifting design combining wheels and vertical elevation. Its vertical working space ranges from 0 to 2 meters. The waist joint's motion range includes ±155° rotation along the Z-axis and movement from -2.5° to +135° along the Y-axis, enabling complex maneuvers such as agile steering.

Furthermore, G1-D offers lower control response latency, with system teleoperation latency under 100 ms and a sampling frequency of 60 Hz. Vertical positioning accuracy is ±0.5 mm, and end-effector gripping precision is ±0.1 mm.

Regarding perception capabilities, G1-D's head is equipped with one high-definition binocular camera, and its hands are fitted with two HD cameras. The Flagship Edition adds LiDAR, depth cameras, physical collision sensors on the chassis, and low-obstacle recognition sensors to reduce collisions in practical use and enhance stability.

Additionally, G1-D's core computing module uses an NVIDIA Jetson Orin NX 16GB, providing 100 TOPS of computational power.

Closing the Loop from Data to Deployment

In addition to the robot itself, Unitree's embodied AI platform behind the G1-D enables rapid capability iteration for robots, covering everything from data collection to model deployment.

At the data collection stage, the platform supports synchronized multi-robot data acquisition. Tasks can be generated with one click via standardized templates, and progress across the entire process is tracked in real time. It is compatible with various robot and end-effector configurations, possesses data standardization capabilities, and can directly output or convert to mainstream training formats. Coupled with a highly available architecture that ensures reliable 24/7 data collection, it significantly reduces data acquisition costs and enhances R&D efficiency.

And in terms of model training and inference tools, it adapts to the open-source ecosystem and deeply integrates open-source models such as PI and GROOT. It supports distributed training and elastic computing power scheduling, with GPU utilization reaching 90%. Users can quickly start model development through the 'one-click training' feature.

Furthermore, the simulation and verification phase relies on a high-fidelity 3D asset library to build realistic simulated environments, enabling the rapid generation of comprehensive algorithm evaluation plans. After model training is completed, it can be seamlessly deployed onto physical robots.

Unitree is accelerating its push into application scenarios

Unitree's capabilities in robot locomotion have been evident to all. They had previously achieved certain results in humanoid robot competitions, but their recent focus has shifted more towards practical scenario implementation.

Today, Unitree unveiled a solution applicable to four major scenarios: service, domestic life, commerce, and industry. Specific tasks include front-desk reception, laundry folding, shelf organization, and inspection.

Recently, Unitree also launched a full-body teleoperation platform. Under this system, the G1 demonstrated practical capabilities in household chores, smoothly executing tasks such as washing dishes, sweeping floors, and serving tea.

From the full-body teleoperation platform to today's G1-D, a key message emerges: Unitree is accelerating its efforts in practical robotic labor.

"First move, then work." This is Unitree's understanding of how robots should be deployed. Now, they have deeply implemented this philosophy.