Unitree G1: Price, Specifications, Development Access, and Safety

The price is real, but incomplete

Unitree has done something important with G1: it has put a specific, orderable price beside a humanoid robot. The standard G1 is listed at $13,500. In a market still crowded with prototypes, pilot programs, private quotes, and futuristic launch videos, that number gives a university, startup, or independent laboratory something unusually concrete to discuss. It makes humanoid hardware feel less like a distant industrial bet and more like equipment that might fit inside a serious research budget.

The number is real, but it is not the cost of a functioning humanoid program. Unitree says tax and shipping are excluded, while its shop estimates freight at $300 to $1,200 and leaves customs duties, local taxes, and import clearance to the buyer. Those are only the visible additions. A team may also need batteries, end effectors, compute, spares, safety barriers, protected floor space, insurance, simulation infrastructure, staff time, travel, training, and a way to recover a 35 kg machine after a fall. The robot is the most photogenic line item. It is rarely the only expensive one.

G1 and G1 EDU are different purchases

The crucial fact is easy to miss because the bodies look similar. Unitree states that the standard G1 does not support secondary development. The development-oriented product is G1 EDU, and its price is not published. Buyers must contact sales. If the purpose of the purchase is custom control, reinforcement learning, imitation learning, manipulation research, or a new embodied-AI stack, this distinction should appear at the top of the procurement brief, not in the footnotes after the robot arrives.

The base G1 has 23 degrees of freedom. G1 EDU can be configured from 23 to 43, depending on waist, wrist, and hand options. EDU also raises the listed maximum knee torque from 90 Nm to 120 Nm, supports secondary development, and can take additional high-compute modules such as NVIDIA Jetson Orin. Unitree lists an eight-month warranty for G1 and an eighteen-month warranty for G1 EDU, subject to the detailed purchase terms. These are not cosmetic package differences. They define what the buyer can build, what hardware can be attached, and how much engineering headroom exists.

A humanoid is not a software subscription

It is tempting to interpret a $13,500 humanoid through the buying habits of software. Pay the entry price, create an account, and begin experimenting. Physical robotics punishes that analogy. Every ambitious feature is attached to mass, torque, power, heat, contact, wear, calibration, cables, and the possibility of a fall. An application that looks trivial in a browser can become a month of integration when it must survive gravity.

G1 folds into a compact shape and weighs about 35 kg with its battery. That is impressively portable by humanoid standards, but it is not casually portable. A lab still needs a safe method for lifting, carrying, storing, charging, and working on it. A fall can damage the robot, the floor, a desk, or a person. Replacement parts and repair time can interrupt research plans in a way that no software restart can solve. Procurement should therefore include an operating model: who owns the robot, who may enable motion, who inspects it, who approves updates, and what happens after an impact.

Agility videos are not deployment evidence

Unitree's demonstrations are effective because they make the machine's control look effortless. Dynamic recovery, fast walking, unusual poses, and whole-body movement communicate real engineering progress. They do not establish that a delivered unit can safely repeat the same behavior in a buyer's building, with a buyer's software, around a buyer's staff. A controlled demonstration is evidence of capability under its conditions. It is not a general warranty of autonomy, reliability, or safety.

That difference matters most when a project proposal jumps from movement to work. Walking across a clear lab floor is not the same problem as navigating a busy workshop. Closing a hand around an object is not the same as reliably picking mixed items for an eight-hour shift. A policy that succeeds in a selected video may still fail under different lighting, friction, battery state, network latency, payload, object geometry, or contact. The honest starting point is not that G1 can already do the job. It is that G1 gives a capable team a physical system on which to investigate whether the job can be done.

The useful machine is the one you can instrument

G1 includes a depth camera and 3D LiDAR, an 8-core CPU, Wi-Fi 6, Bluetooth 5.2, four microphones, a speaker, and a manual controller. Those components create a credible base for perception, navigation, interaction, and supervised testing. On EDU configurations, optional compute and development access make the platform much more interesting for teams building their own stack.

Yet a specification sheet is not an experiment design. Researchers need timestamps, calibration records, synchronized logs, repeatable software images, and a clear account of which processes run on the robot and which run elsewhere. They need to know what happens when a sensor stream stalls, wireless connectivity drops, a controller misses its deadline, or a model produces a confident bad action. Instrumentation is what turns an impressive movement into knowledge. Without it, the lab is collecting anecdotes.

Hands change the research question

The standard G1 has no hands. G1 EDU can be ordered with optional grippers or Unitree's Dex3-1 three-finger hands, with optional tactile sensor arrays and additional wrist motion depending on the configuration. This is more than an accessory choice. The end effector determines what data can be collected, what policies can be trained, what objects can be handled, and which failures become likely.

A gripper can be the better scientific instrument when the goal is reliable pick and place. A dexterous hand makes sense when the research question concerns contact-rich manipulation, hand-object coordination, tactile learning, or human-like demonstrations. It also adds joints, sensing, calibration, fragile contact points, policy complexity, and cost. Teams should resist paying for anthropomorphic elegance without a task that needs it. The most useful hand is the one whose geometry, force, sensing, and failure modes match the object set.

The same caution applies to payload. Unitree gives approximate arm-load figures of about 2 kg for G1 and 3 kg for G1 EDU, while warning that capacity varies greatly with arm extension and posture. That caveat is the meaningful part. A mass held close to the torso at low speed is a different problem from the same mass extended forward while the robot turns. Buyers should validate the complete motion envelope, not copy a single kilogram number into a requirements table.

Two hours is a planning number

Unitree advertises about two hours from the quick-release 9000 mAh battery. That is useful for planning, but it is not a guaranteed two-hour experiment. Dynamic walking, repeated recovery, onboard AI compute, hand actuation, wireless traffic, payload, temperature, and battery age all shape runtime. So does the amount of time the robot spends standing while researchers inspect logs and revise code.

A good lab measures its own energy profile early. Record battery state, voltage, temperature, compute load, motion state, and shutdown margin across representative sessions. Decide whether a second battery improves utilization, how batteries will be stored, who checks their condition, and how a low-energy event affects balance and safe shutdown. Quick release can shorten downtime, but only if battery handling is treated as an operating process rather than an afterthought.

The lab is part of the robot

The most mature G1 deployment may look less dramatic than the videos. It begins with a marked test area, a clear fall zone, protected equipment, low speeds, conservative joint limits, a remote stop procedure, and an operator who is not also trying to film. Access is controlled. The floor is checked. Loose cables disappear. Someone is assigned to stop the test, and everyone knows that person's authority is immediate.

This is not timidity. It is how a laboratory preserves both people and hardware long enough to learn. A roughly 35 kg humanoid carries meaningful energy even before it lifts an object. Pinch points, foot placement, unexpected recovery motions, balance loss, battery events, and software faults deserve explicit controls. Public demonstrations need stronger controls still, because spectators do not know the robot's boundaries and may step into them.

The same discipline belongs on the network. G1 includes cameras, microphones, Wi-Fi, Bluetooth, remote-control functions, software updates, and developer interfaces. A serious deployment should segment the robot network, protect credentials, limit developer access, record approved versions, review third-party packages, and define how logs and captured media are stored. Firmware and over-the-air updates should be tested before a scheduled demonstration or experiment. An update that improves one behavior can still alter timing, dependencies, or reproducibility elsewhere.

Open source improves the starting line

Unitree's software ecosystem is one of G1's strongest arguments. The company publishes SDK2 Python support, G1 assets and workflows for Unitree MuJoCo and Unitree RL Gym, manipulation datasets, LeRobot-oriented imitation-learning resources, and work around its UnifoLM embodied-intelligence stack. These resources do not make humanoid development easy, but they reduce the blank-page problem. Teams can begin with recognizable simulation, control, and learning workflows rather than inventing every interface themselves.

The practical advantage is strongest when a lab uses the ecosystem as a reproducible baseline. Pin repository commits. Record licenses. Archive environment files and robot configurations. Keep simulation and hardware logs connected to the model that produced them. Document calibration, sensor layouts, joint options, firmware, and end effectors. A research result that cannot be tied to a complete system state is difficult to reproduce and harder to trust.

Open code also does not erase the gap between simulation and hardware. Contacts differ. Friction changes. Joints have backlash and wear. Sensors are noisy. Timing is imperfect. Policies should be trained and stressed with variation, then transferred at reduced speed and range in a controlled area. Each increase in energy, reach, payload, or autonomy should be earned by evidence from the previous stage.

The procurement meeting should start with failure

The strongest G1 business case is not the one with the most spectacular target behavior. It is the one that can explain what happens when that behavior fails. Who pays for a damaged hand? How many falls can the schedule absorb? Is a spare battery enough, or is a second robot needed to protect a course or study from downtime? Can the team diagnose a network failure separately from a controller failure? Does the warranty cover the anticipated use, and where will repairs happen?

A quote for G1 EDU should specify the exact degree-of-freedom configuration, knee and arm hardware, compute, end effectors, tactile options, accessories, batteries, charger, documentation, software access, support, training, warranty, delivery terms, and replacement-part process. The team should also request clarity on any interface or behavior that is essential to the project. Verbal assumptions are cheap before a purchase and costly after it.

This is where the base G1 can still be the right answer. A teaching program focused on observation, supported motion, system identification, perception, or supervised demonstrations may not need the deeper development surface. The mistake is not buying the smaller package. The mistake is buying it for objectives that require the larger one.

The alternatives are moving fast

G1 no longer sits alone as the obvious low-price humanoid story. Unitree now markets the smaller R1 at a lower starting price, while its H2 family moves toward larger and more capable embodiments at higher prices. Booster T1 is another compact research-oriented option. Fourier GR-1 targets a larger platform, while systems such as Apptronik Apollo approach humanoids from an industrial deployment direction. These are not interchangeable products, and some alternatives may be available only through programs, partnerships, or private quotes.

Comparison should begin with the job and the access model. A lower list price is irrelevant if the required development interfaces are absent. A larger payload is not automatically better if the robot is harder to house safely. A polished industrial pilot may offer more support but less experimental freedom. A research platform may expose more of the stack while leaving integration and reliability to the buyer. Availability, delivery time, configuration transparency, support location, software rights, and replacement parts can decide the outcome before locomotion specifications do.

For many laboratories, G1 EDU's attraction is the combination of a compact body, a recognizable developer ecosystem, official simulation resources, optional manipulation hardware, and a company willing to sell the platform. That is a meaningful combination. It should still be compared against the actual project, not against the emotional pull of owning a humanoid.

The WhatAI verdict

Unitree G1 makes humanoid robotics more accessible, but accessibility is not the same as simplicity. The $13,500 standard model is a real product at a striking price. It is also the version without secondary development. The machine most serious robotics teams are imagining when they say they want a G1 is often G1 EDU, configured with the right joints, compute, and end effector, then surrounded by a capable lab.

That does not weaken the case for G1. It clarifies it. This is a promising platform for universities, robotics researchers, and embodied-AI teams that can define a bounded question, buy the correct configuration, begin in simulation, transfer cautiously, and measure what the robot actually does. Unitree's software resources make the starting line better than it once was, and the compact hardware creates room for experiments that used to demand far larger budgets.

The wrong buyer sees a $13,500 autonomous worker. The right buyer sees a $13,500 entry point, an EDU decision, a stack of optional hardware, a safety program, and months of careful engineering. One is a product fantasy. The other is a credible research plan.

WhatAI Decision Box

Unitree G1 is a compact humanoid robot starting at $13,500. The standard 23-DoF G1 and development-oriented G1 EDU share the same basic form, but they serve different buyers. Secondary development, expanded joint configurations, optional high compute, and dexterous end effectors are G1 EDU territory.

The headline price excludes shipping, tax, customs duties, and import clearance. A working research program may also need the EDU model, hands or grippers, compute, spare parts, simulation, safety equipment, and experienced staff. Unitree lists about two hours of battery life, but laboratories should validate runtime, payload, balance, thermal behavior, and software versions against their own tasks.

G1 is strongest as a supervised research, teaching, and prototyping platform. Choose the base model for supported demonstrations and limited experimentation. Choose G1 EDU when custom software, low-level development, expanded degrees of freedom, or manipulation hardware are essential. Neither should be treated as a general-purpose autonomous worker without substantial engineering and safety controls.

Use Cases

Pricing varies by plan and region — see current pricing.

Plan features change — last updated: 2026-08-16.

Unitree G1 Community Discussions

Explore community discussions. Ask and answer questions on Unitree G1 to grow and learn together.

Unitree G1 Pros & Cons

How to Get Results with Unitree G1: Step-by-Step Workflow

1. Define the research outcome

   Specify the behavior, environment, payload, human proximity, success criteria, data needs, control level, and acceptable failure modes before selecting a configuration.

2. Choose G1 or G1 EDU

   Use the standard G1 only when supported behaviors and its closed development scope are sufficient. Select G1 EDU for secondary development, optional hands, expanded joints, or higher compute, and obtain the complete quote in writing.

3. Build the total-cost budget

   Add freight, duties, tax, insurance, end effectors, compute, batteries, chargers, spares, service, safety equipment, simulation, staff time, travel, and training to the robot price.

4. Design the safety cell

   Create a controlled test area with fall clearance, physical exclusion, protected equipment, remote stop procedures, access rules, spotters, lifting plans, and documented recovery steps.

5. Freeze the software baseline

   Record robot configuration, firmware, SDK, dependencies, compute image, network settings, credentials, licenses, and calibration state. Test updates in isolation before promoting them to the main robot.

6. Prototype in simulation

   Develop and stress policies in Unitree MuJoCo, Unitree RL Gym, or another validated simulator. Add observation noise, delays, contact variation, disturbances, and conservative joint and torque limits.

7. Commission at low energy

   Verify communications, controls, emergency procedures, joint limits, sensor streams, calibration, battery behavior, logging, and safe shutdown before enabling dynamic motion or carrying objects.

8. Transfer one capability

   Move a bounded behavior from simulation to hardware with reduced speed and range. Use spotters and an exclusion zone, measure tracking and contact errors, and stop on predefined thresholds.

9. Measure the real envelope

   Benchmark runtime, thermals, balance, payload by posture, network loss, recovery, repeatability, floor sensitivity, sensor limitations, and degradation across repeated sessions.

10. Operate with change control

    Log every run, inspect the robot after falls, manage batteries, back up configurations, review firmware updates, rotate credentials, track component wear, and revalidate behavior after any material change.

Unitree G1 Gotchas and Limits to Know Before You Start

• The $13,500 price applies to the standard G1 and excludes tax and shipping.
• The base G1 does not support secondary development; G1 EDU does.
• G1 EDU pricing is not publicly listed and can rise with joints, compute, hands, and support.
• Official shipping estimates exclude customs duties, taxes, and import clearance.
• The base G1 has 23 degrees of freedom, while G1 EDU ranges from 23 to 43.
• Dexterous hands, tactile arrays, grippers, wrist joints, and high compute are configuration-dependent options.
• The approximate arm-load figures depend strongly on reach, posture, motion, and balance.
• About two hours of battery life is an advertised figure, not a guaranteed workload result.
• A roughly 35 kg humanoid can injure people or damage itself and nearby equipment when it falls.
• Dynamic marketing demonstrations do not prove autonomous task reliability or safe operation.
• Over-the-air updates can alter behavior, dependencies, or compatibility and require regression testing.
• Camera, microphone, Wi-Fi, and Bluetooth use create privacy and cybersecurity obligations.
• Sim-to-real transfer still needs conservative limits, controlled testing, data logging, and repeated validation.
• Warranty length and coverage differ by model and should be confirmed in the purchase documents.

Which Unitree G1 Feature Fits Your Use Case

Starter Prompts for Unitree G1

How much does Unitree G1 cost?

Unitree lists the standard G1 at $13,500 before tax and shipping. Its official shop estimates shipping at $300 to $1,200 and says customs duties, taxes, and import clearance remain the buyer's responsibility. G1 EDU pricing requires a sales inquiry.

What is the difference between G1 and G1 EDU?

The standard G1 has 23 degrees of freedom and does not support secondary development. G1 EDU supports secondary development, 23 to 43 degrees of freedom, higher knee torque, optional high-compute modules, optional end effectors, and a longer listed warranty.

Can developers run custom software on Unitree G1?

Unitree states that the base G1 does not support secondary development. Buyers who need custom development should select G1 EDU and confirm the exact hardware, interfaces, permissions, and support scope with Unitree before ordering.

Does Unitree G1 include dexterous hands?

No. The base G1 has no hands. G1 EDU can be configured with optional grippers or Dex3-1 three-finger hands, with optional tactile sensor arrays and wrist degrees of freedom depending on configuration.

How long does the G1 battery last?

Unitree advertises about two hours from the quick-release 9000 mAh battery. Actual runtime will depend on motion, payload, compute, accessories, temperature, battery condition, and idle time, so teams should measure it under their own workload.

Is Unitree G1 safe to use around people?

G1 is a roughly 35 kg dynamic machine capable of falling and moving unexpectedly. It should begin in a controlled area with exclusion zones, trained operators, an emergency-stop plan, low-energy tests, fall clearance, and task-specific risk assessment. Marketing demonstrations are not safety validation.

What software resources support G1?

Unitree publishes SDK2 and SDK2 Python resources, G1 support in Unitree RL Gym and Unitree MuJoCo, manipulation datasets, LeRobot-oriented imitation-learning work, and UnifoLM resources. Compatibility and license terms should be verified for the exact project and robot configuration.

Is Unitree G1 a turnkey autonomous worker?

No. It is best understood as a humanoid hardware and research platform. Useful autonomous behavior still requires task design, perception, control, data, testing, supervision, safety engineering, maintenance, and integration with the surrounding environment.

Related Humanoid Robots Tools

Explore the Network