High-performance humanoid robot
Multiple humanoid body options featuring in-house actuators, gear reducers, encoders, and sensors.
UNITREE · Humanoid robots
Unitree G1-D
End-to-end platform for humanoid robots: data and training
An integrated platform spanning robot hardware, data collection, model training, and deployment on real robots.
Core components
The G1-D solution brings the robot body, data acquisition tools, and model training/inference infrastructure together on a single platform.
Streamlined data acquisition tools
A unified data pipeline covering acquisition, processing, labeling, review, and data asset management.
Comprehensive model training and inference tools
Distributed training, custom model development and deployment; support for popular open-source frameworks.
Higher-DOF platform, wide workspace, and low-latency control
The manufacturer lists a total of 19 degrees of freedom (excluding the end effector): arms 7×2, waist 2, column 1, and base 2 DOF. The mobile lifting design combines a wheeled base with vertical movement; the vertical workspace is 0–2 m. The waist range of motion is ±155° on the Z axis and −2,5° to +135° on the Y axis. Lifting accuracy is listed as ±0,5 mm, end-effector gripper accuracy as ±0,1 mm, teleoperation latency as less than 100 ms, and the sampling rate as 60 Hz. Accuracy may vary depending on the end-effector configuration.
19
Robot DOF
Total degrees of freedom, excluding the end effector.
7×2
Arm DOF
Combined total for both arms; 7 DOF per arm.
2 / 1 / 2
Waist / column / base DOF
Waist 2, column 1, and base 2 degrees of freedom.
0–2 m
Vertical workspace
Vertical reach provided by the mobile lifting design.
<100 ms; 60 Hz
Control
Teleoperation latency and sampling rate.
Accuracy values vary depending on the end-effector configuration. This capability overview describes the 19 DOF configuration; the table below lists separate values for the Standard variant.
Streamlined data acquisition tools
Aims to simplify data collection and reduce costs through flexible, standardized workflows, moving beyond traditional manual processes. Visual template management combines project management, task assignment, progress tracking, and status analysis; tasks can be created from a template with one click. Multiple robot platforms and end effectors are supported; standardization aims to process data from various devices end to end into high-quality data for model training. The high-concurrency architecture is described as supporting simultaneous data collection by hundreds of robots. The high-availability service architecture aims to enable 24/7 data collection and convert data into widely used training formats.
Centralized task management
Visual template management for efficient collection
Projects, task assignments, progress, and status analysis in one interface; create a task from a template with one click.
Multi-platform
Flexible platform and end-effector configuration
Standardizes data from various robots and end effectors and converts it into training data.
Hundreds of robots
Scalable high concurrency
Described as capable of receiving and processing data streams from hundreds of robots through load balancing and concurrent scheduling.
Continuous collection objective
24/7 online collection
High availability, format compatibility, and output/conversion to widely used training formats.
Presented by the manufacturer as a platform capability; these are not service-level commitments made by Robotik Çözümler.
Data acquisition pipeline
The six-stage data collection workflow listed by the manufacturer.
1
Create a new collection task
Start a new data collection job.
2
Edit and assign the task
Edit task settings and assign it to a responsible person.
3
Data acquisition and annotation
Collect and label robot data.
4
Upload and review data
Upload and review the collected data.
5
Data storage
Store approved data.
6
Export data
Export data for subsequent use.
Comprehensive model training and inference tools
Provides an end-to-end workflow from data processing to one-click model deployment and integrates with widely used open-source robotics model frameworks. Describes integration with open-source models such as PI and GROOT, training with community/open datasets, and a high-accuracy 3D asset library and simulation environment for algorithm validation. Aims for out-of-the-box deployment, one-click training, and sim-to-real transfer. The distributed training infrastructure flexibly schedules compute tasks and accelerates them in parallel; the manufacturer states that GPU utilization can reach up to %90.
Open model ecosystem
Ecosystem supporting widely used models
Community and open datasets; deep integration with models such as PI and GROOT.
Simulation
Model evaluation simulation
Algorithm validation scenarios with high-accuracy 3D assets and realistic environments.
One-click training
User-focused design for rapid deployment
Aims for one-click training, integrated simulation, and transfer from algorithm to real machine.
GPU utilization: up to %90 (manufacturer claim)
Efficient distributed training
Flexible task scheduling and parallel acceleration; scales according to resource load.
The %90 GPU utilization figure for Unitree G1-D is a claim on the manufacturer's page; it is not an independent performance guarantee.
Training and inference pipeline
Model development, evaluation, and deployment sequence.
1
Select model architecture
Determine the model structure to be used.
2
Training configuration
Set training parameters.
3
Real-time monitoring
Track the training process.
4
Parameter editing
Edit model parameters.
5
Simulation testing
Test the model in a simulated environment.
6
Export and deploy the model
Export the trained model and put it into deployment.
Application scenarios
The manufacturer's page lists application areas under four headings.
Service
Scenario shown for service applications.
Life
Scenario for living environments.
Retail
Retail application scenario.
Industry
Industrial application scenario.
UnifoLM-WMA-0: World Model-Action (WMA) Framework
UnifoLM-WMA-0 is presented as Unitree's open-source world model-action architecture, designed for general-purpose robot learning and covering different robot embodiments. Its core world model aims to understand physical interactions between a robot and its environment. In decision mode, it predicts future interactions based on the environmental state and task goals; these predictions assist the action policy, aiming to reduce decision errors and improve movement accuracy. In simulation mode, it generates high-accuracy environmental feedback in response to robot motion inputs; synthetic data supports model training and policy optimization.
Decision-making
Decision mode: action generation through prediction
Predicts physical interactions based on the state and task goals, helping the policy module generate actions.
Simulation
Simulation mode: high-accuracy feedback for data generation
Supports the training/optimization process by generating environmental feedback and synthetic data in response to motion input.
Product parameters
The manufacturer compares the G1-D Standard and G1-D Ultimate variants. The table should be retained with its headings for mechanical dimensions, base, electrical specifications, accessories, and other items.
| Variants | G1-D Standard | G1-D Ultimate |
|---|---|
| Overall dimensions, column at minimum height | 1260×620×620 mm | 1260×525×570 mm |
| Overall dimensions, column at max. height | 1680×620×620 mm | 1680×525×570 mm |
| Total weight (including battery) | 90 kg | 90 kg |
| Total DOF (excluding end effector) | 17 | 19 |
| Single-arm DOF (excluding end effector) | 7 | 7 |
| Maximum payload per arm [1] | Approx. 3 kg | Approx. 3 kg |
| End-effector options [2] | Same options | Same options |
| Waist DOF / joint range of motion | 2 / same range | 2 / same range |
| Column lifting speed | 60 mm/s | 60 mm/s |
| Maximum movement speed | Not specified (/) | 1,5 m/s |
| Chassis drive type | Not specified (/) | Differential drive |
| Chassis sensors | Not specified (/) | LiDAR ×1 + depth camera ×2 + collision sensor ×2 + low-obstacle sensor ×2 |
| Base computing power | 8-core CPU | 8-core CPU |
| Perception sensors | Same | Same |
| Wi‑Fi 6, Bluetooth 5.2 | Yes | Yes |
| High-computing-power module | Various options | Various options |
| Battery | Upper body 9 Ah + chassis 30 Ah |
| Manual control unit / visualization computer | Yes / Yes | Yes / Yes |
| Battery operating time | Approx. 2 hours | Approx. 6 hours |
| Smart OTA upgrade / secondary development [3] | Yes / Yes | Yes / Yes |
[1] The arm's maximum payload can vary significantly depending on the arm extension posture. [2] End-effector selection depends on the configuration. [3] Refer to the manufacturer's development guide for details on secondary development. The manufacturer also states that parameters may vary by scenario/configuration, that a safe distance should be maintained because of the robot's powerful and complex design, that the appearance may change, and that some sample functions are in the development/testing stage. Do not present any warranty or sales terms here as a company promise.
Safety and product notes
The manufacturer describes this product as a civilian-use robot; it requires that hazardous modifications or use be avoided, local laws be followed, and a sufficient safe distance be maintained between the robot and people. The product appearance and parameters may change; some sample functions may be in the development and testing stage.
This is a manufacturer statement and safety warning; it is not a warranty or performance commitment from Robotik Çözümler.
Next steps for Unitree G1-D.
Let's evaluate your model selection and application needs together.