Unitree G1 vs H1: Which Humanoid Should Developers Choose?

The Unitree G1 and H1 can both walk, use Unitree SDK 2, and serve as embodied AI research hardware. They are not interchangeable. G1 is a compact, lower-cost platform with a growing learning and simulation ecosystem. H1 is a full-size locomotion platform, while H1-2 adds a more useful manipulation configuration. The larger machines demand more space, stronger safety controls, and a sales-led budget.

For most university labs and developer teams entering humanoid research, G1 EDU is the practical choice. H1 or H1-2 makes sense when human-scale reach, stronger joints, or full-size locomotion is central to the research question. The $13,500 base G1 should not be mistaken for the programmable G1 EDU.

This comparison uses Unitree’s official specifications checked on August 20, 2026. Configuration and regional availability can change, so request a written bill of materials before purchasing.

For the wider portfolio and ecosystem context behind this decision, start with Unitree Robots Explained.

G1 vs H1 quick comparison

Attribute	G1 base	G1 EDU	H1	H1-2
Position	Demonstration-oriented compact humanoid	Compact research humanoid	Full-size locomotion platform	Full-size locomotion and manipulation platform
Height	1.32 m	1.32 m	About 1.80 m	About 1.78 m
Weight	About 35 kg	About 35 kg or more	About 47 kg	About 70 kg
Total joints	23	23 to 43	Core configuration varies by expansion	27
Arm joints	5 per arm	5 per arm, optional wrists/hands	4 per arm, expandable	7 per arm
Sensors	Depth camera and 3D LiDAR	Depth camera and 3D LiDAR	Depth camera and 3D LiDAR	Depth camera and 3D LiDAR
Battery	9,000 mAh, about 2 hours claimed	Same core specification	864 Wh, quick-release	864 Wh, quick-release
User compute	Basic 8-core CPU	Optional high-compute modules including Orin-class options	Intel i7 user-development computer listed	Intel i7 plus optional compute, up to three devices listed
Secondary development	Not included in official table	Yes	Developer interfaces available	Developer interfaces available
Price	$13,500 before tax and shipping	Contact sales	Contact sales	Contact sales
Best fit	Exhibitions and packaged motions	Learning, data collection, VLA and control research	Dynamic full-size locomotion	Full-size manipulation and whole-body research

The H1 page provides specifications rather than a current public checkout price. Historical listings or reseller quotes can help estimate budget, but they are not a substitute for a current Unitree quotation. See the living humanoid robot price comparison for price labels and verification dates.

Price and total project cost

G1’s $13,500 headline makes the platform look like a simple purchase decision. It is only the base robot price, excluding tax and shipping, and Unitree does not mark that base configuration for secondary development. A research team needs a G1 EDU quote.

The quote should specify the joint configuration, hands or grippers, compute module, batteries, charger, controller, support, warranty, shipping, duties, and included software access. The official G1 page lists 23 to 43 joints for EDU because optional waist joints, wrist joints, and seven-joint Dex3-1 hands change the machine substantially.

H1 and H1-2 require contact with sales. Their total cost also includes facility changes that a compact G1 project may avoid. A 70 kg H1-2 with powerful joints needs a controlled test area, lifting or suspension equipment, trained operators, stronger transport cases, and potentially more expensive spare assemblies.

Budget for the complete experiment:

• robot and selected configuration;
• shipping, duties, tax, and insurance;
• at least one spare battery when continuous experiments matter;
• optional hands and their replacement parts;
• host compute, networking, and data storage;
• teleoperation hardware;
• safety barriers, emergency controls, and suspension equipment;
• spare joints, cables, protective parts, and tools;
• staff time for calibration, integration, and recovery.

The robotics development kit guide offers smaller alternatives for teams that mainly need to learn imitation learning or manipulation. A low-cost arm can answer many software questions before a humanoid purchase.

Size, reach, and workspace

G1 is easier to fit into a lab. At 1.32 meters and roughly 35 kilograms, it can work with smaller safety zones and is easier to transport than a full-size humanoid. It is still powerful enough to cause injury or damage, so compact does not mean safe for unsupervised use.

H1 is closer to adult human height but relatively light at about 47 kilograms. Its published 3.3 meter-per-second moving speed demonstrates dynamic locomotion. That number should not drive a buying decision unless high-speed locomotion is the research topic. Most manipulation research values controllability, reach, sensing, and recovery over top speed.

H1-2 weighs about 70 kilograms and provides a more complete upper body. Unitree lists seven degrees of freedom per arm and six per leg, plus optional Dex5-1 or other hands. That creates better reach and wrist orientation for human-scale workstations. It also increases collision energy, control complexity, and the cost of falls.

Choose the smallest embodiment that can represent the task. If the project studies tabletop manipulation, learning pipelines, or VLA adaptation, G1 may be enough. If it studies industrial totes at multiple heights, human-scale reach, or full-body load transfer, H1-2 is more relevant.

Degrees of freedom and manipulation

Joint count is often used as a capability score, but it only describes available motion axes. It does not measure control quality, hand reliability, perception, policy performance, or task success.

The base G1 has 23 joints. EDU configurations can add waist freedom, wrists, and dexterous hands, reaching 43 joint motors. Unitree lists about 3 kg maximum arm load for G1 EDU, with an explicit note that load varies with arm extension. That is not the same as a rated payload across the entire workspace.

H1’s original configuration prioritizes locomotion. Four-joint arms are less suitable for general manipulation because wrist pose and approach angle are constrained. H1-2 addresses that limitation with seven-joint arms. Unitree lists an H1-2 arm normal load of about 7 kg and a peak value around 21 kg. A procurement test should use the required reach, speed, wrist pose, and duty cycle rather than the peak figure.

Hands are separate systems. A dexterous hand adds actuators, tactile sensing options, calibration, grasp planning, and failure modes. It can be necessary for research, but a simple gripper is often more reliable for a narrow task. Ask whether the hand is included, which SDK controls it, and whether simulation and spare fingertips are available.

Sensors and onboard computing

Both families list 3D LiDAR and a depth camera. Those sensors support mapping, obstacle perception, and scene capture, but the product page does not establish performance for every environment. Reflective surfaces, sunlight, dark materials, motion blur, and occlusion can degrade depth sensing.

G1 includes a basic eight-core CPU, while EDU offers optional higher-compute modules, including Orin-class choices. H1 lists separate platform and user-development computers, with optional Intel or Orin configurations. H1-2 lists space for additional compute.

Do not select a robot based only on TOPS or CPU names. Start from model size, precision, camera rate, preprocessing, policy frequency, thermal envelope, and latency budget. A VLA can reason at one rate while a lower-level controller runs much faster. The VLA models guide explains these model roles.

SDK, ROS 2, simulation, and learning workflows

Unitree SDK 2 supports G1, H1, and H1-2 through DDS-based interfaces. The public repository contains high-level services and low-level command examples. Python bindings and ROS 2 packages are available separately. Unitree also publishes MuJoCo and Isaac Lab resources.

The ecosystem is strongest around G1 for current learning workflows. Unitree’s repositories support XR teleoperation for 23 and 29 degree-of-freedom G1 variants. LeRobot documentation and Unitree integrations cover data collection and policy training. NVIDIA provides an end-to-end GR00T workflow for G1.

H1 and H1-2 are supported in SDK and teleoperation repositories, but teams should verify the exact branch, joint mapping, hand, and simulator assets for their purchased revision. H1-2’s parallel ankle mechanism has its own control considerations documented in the official SDK examples.

A sensible workflow is:

1. Verify the exact robot revision and joint map.
2. Bring up high-level control while the robot is secured.
3. Validate emergency stop, limits, and mode transitions.
4. Reproduce the robot in simulation.
5. Collect task data through teleoperation.
6. Train or adapt a policy offline.
7. Test deterministic cases and failure conditions in simulation.
8. Deploy at reduced speed and torque under supervision.
9. Log interventions, near misses, and recovery events.
10. Roll back immediately when predefined limits are exceeded.

The sim-to-real robotics workflow provides the full validation structure. The robot API and SDK directory explains how Unitree tooling fits with ROS 2, Isaac, LeRobot, and other components.

Which robot fits which developer?

Choose G1 EDU for accessible humanoid research

G1 EDU is the default recommendation for universities, robotics courses, embodied AI startups, and model researchers. It costs less, needs less space, and has strong documentation across third-party learning stacks. It is suitable for locomotion, imitation learning, compact manipulation, human-robot interaction, and whole-body policy research.

Do not choose the base G1 when custom development is the objective. Do not assume a 23-joint EDU quote includes the hands or wrist control needed for manipulation.

Choose H1 for dynamic locomotion research

H1 is appropriate when full-size biped dynamics, speed, disturbance recovery, or powerful leg actuation is the research target. Its original arm configuration is less attractive for general manipulation. Buyers should compare it with H1-2 and newer H2 options before ordering.

Choose H1-2 for full-size manipulation research

H1-2 is the more logical choice when human-scale reach, stronger arms, and wrist orientation matter. It can represent industrial workstation geometry better than G1. The tradeoff is greater facility, safety, maintenance, and integration burden.

Choose neither when the research question is smaller

A humanoid is unnecessary for many manipulation projects. A fixed arm is easier to secure, instrument, reset, and reproduce. A quadruped is more suitable for mobility and inspection. Simulation is appropriate for early architecture and control work. The cheapest robot is the one that answers the actual research question without adding unused complexity.

Limitations shared by G1 and H1

Neither family is delivered as a general autonomous worker. Built-in motions establish mobility, not task generality. Custom work requires a development-capable configuration, data, a policy, application logic, safety controls, and repeated validation.

Battery duration limits continuous operation. Falls can damage expensive components. Dexterous manipulation remains sensitive to calibration and object variation. Vendor videos do not report intervention rate, recovery time, task completion distribution, or maintenance hours.

Researchers should define measurable acceptance criteria: successful cycles, human interventions, dropped objects, unsafe contacts, recovery time, energy per task, and hours between maintenance events. Compare systems on the same task and protocol rather than using unrelated vendor demonstrations.

Decision examples

A university teaching embodied AI: choose G1 EDU with a documented joint configuration, spare battery, and simple gripper or hand appropriate to the course. Its smaller workspace and public learning integrations matter more than H1’s peak locomotion performance.

A lab studying dynamic biped locomotion: compare H1 with H1-2 and newer full-size platforms. H1’s lower mass and published high-speed mobility may be relevant, but the team should obtain the exact controller, suspension, and simulation configuration used for its experiments.

A team studying warehouse manipulation: start from reach, tote height, payload, and gripper requirements. H1-2 may represent the workstation better than G1, but a wheeled manipulator or fixed arm may provide higher uptime and simpler safety. Run the economic comparison before choosing a humanoid form.

A foundation-model startup: G1 EDU is usually the lower-cost embodiment for data and policy work. Buy identical configurations if multiple robots will share datasets. Hardware variation can become an uncontrolled training variable.

An exhibition buyer: the base G1 may provide packaged motions without a research configuration. Confirm which behaviors are licensed and supported, how the robot is supervised, and whether the venue can maintain a safe operating perimeter.

In every scenario, require the quotation and acceptance plan to name the exact robot revision. G1, G1 EDU, H1, and H1-2 are not software-compatible labels by themselves. Joint maps, hands, compute modules, firmware, and simulator models determine whether a published workflow can be reproduced on the delivered machine.

Verdict

Choose G1 EDU for most developer and university projects. It is the lower-risk way to access a real humanoid, and its integration with Unitree SDK 2, ROS 2, LeRobot, Isaac, and public simulation resources creates a useful learning path.

Its compact scale also makes repeated supervised experiments, transport between workspaces, and classroom access more manageable than with a full-size platform.

Choose H1 only when its full-size locomotion is the point. Choose H1-2 when human-scale manipulation and stronger arms justify the additional cost and safety burden. Before purchasing either, compare newer H2 configurations and request written confirmation of the exact software rights, compute, hands, warranty, and support.

For configuration detail, read the Unitree G1 complete guide. For market context, see Unitree versus Figure and Tesla.

Frequently asked questions

Is the $13,500 Unitree G1 programmable?

Unitree’s official table does not mark the base G1 for secondary development. Researchers should request G1 EDU and confirm interfaces in writing.

Is H1 better than G1?

Not generally. H1 is larger and more powerful. G1 is cheaper, easier to house, and has a strong learning ecosystem. The better platform depends on the experiment.

Does H1-2 replace H1?

H1-2 improves full-size manipulation with seven-joint arms and revised legs, but H1 remains relevant for locomotion work. Availability and configuration should be confirmed with Unitree.

Can G1 run robot foundation models?

Yes, with the correct EDU configuration, compute, interfaces, and adaptation work. NVIDIA and Unitree publish relevant workflows. A checkpoint still needs task data, action mapping, safety limits, and validation. See the robot foundation model guide.

Should students buy a humanoid or a robot arm?

A robot arm is usually better for learning manipulation on a limited budget. Choose a humanoid only when biped locomotion, whole-body control, or humanoid embodiment is central to the project.

Sources

• Unitree G1 specifications - dimensions, joints, sensors, battery, development and price
• Unitree H1 and H1-2 specifications - dimensions, joint layouts, compute and payload claims
• Unitree SDK 2 - official control examples and supported robots
• Unitree XR teleoperation - supported G1 and H1 configurations
• Unitree MuJoCo - official simulation resources
• NVIDIA GR00T workflow for G1 - model adaptation and deployment workflow