Unitree G1 Review: The Cheap Humanoid's Expensive Catch

The Unitree G1 did something no humanoid had managed before: it put a walking, articulated, commercially available bipedal robot inside the price band of a used car. That single fact is why it dominates the search results, the conference floors and the YouTube demo reels. It is also why so many buyers end up with the wrong unit, because the headline price and the machine most buyers actually need are separated by roughly a factor of three and a hardware boundary you cannot cross after the fact.

Unitree revises pricing often, so treat every figure below as a checked-on-the-day reading rather than a standing quote. The structure is what lasts: an entry tier, a much more expensive developer tier, and no bridge between them.

This review is about that gap, plus the two things most G1 coverage leaves out: what the runtime really is, and the unpatched security flaw that has been public since September 2025.

Companion reading: how to choose a humanoid robot, humanoid robot hardware, robot cybersecurity, legged and quadruped robot hardware, and imitation learning for robotics.

Table of contents

1. Key takeaways
2. The price ladder, and the trap in it
3. What the hardware actually is
4. Runtime, payload, and the physics of a small humanoid
5. The software stack
6. The security problem nobody demos
7. How it compares
8. Who should buy which tier
9. Verdict
10. Frequently asked questions
11. Changelog

The price ladder, and the trap in it

As of August 2026, Unitree listed the G1 in the low-to-mid teens of thousands of US dollars on its own site, with resellers landing roughly 20% to 60% higher once shipping and import are added. Those specific numbers will move. For a walking humanoid, an entry price in that band is still the reason the category suddenly has a mass market.

Then you read the development terms. The base unit ships with 23 degrees of freedom and fixed gripper hands, and it does not support secondary development. You can operate it. You cannot write software for it.

All figures in thousands of US dollars, read in August 2026. Check the current quote before you budget: the ratio between the tiers has been more stable than either number.

The detail that costs institutions real money: the tiers are not upgrade paths. Motor mounts, wiring harnesses and compute module slots differ between Basic and EDU, so a university that buys three Basic units for a teaching lab and then wants a research platform is buying three more robots. Decide the tier on the first purchase order, because the second one is not a discount.

If you are still choosing a category rather than a model, our how to choose a humanoid robot guide covers segments and specs before brands enter the picture.

What the hardware actually is

Strip the marketing and the G1 is a small bipedal platform, roughly child-height, built around Unitree's own actuator line. The EDU configurations raise joint torque from about 90 N·m to 120 N·m, which is the difference between a robot that can demonstrate a motion and one that can hold a load through it.

Degrees of freedom scale with tier, from 23 on Basic to 43 on the top EDU configurations. Most of that delta is in the hands and wrists, which is exactly where humanoid usefulness lives. A 23-DoF G1 with fixed grippers can walk, gesture and pick simple objects. It cannot do the dexterous manipulation the category is sold on.

Compute is the other tier boundary. EDU units carry an NVIDIA Jetson Orin at around 100 TOPS, enough to run perception and a learned policy on-board. For what that class of compute does and does not buy you, see edge AI and robot compute.

Runtime, payload, and the physics of a small humanoid

Two numbers govern what you can actually do with a G1, and both are smaller than the marketing suggests.

Runtime is about two hours of active use. Figures in the five-to-ten-hour range describe standby, not walking and manipulating. A humanoid spends its energy budget keeping itself upright, and a small battery in a small body does not change that. Any deployment plan that assumes a shift needs either hot-swap batteries and a person to swap them, or a docking routine and the acceptance that the robot is offline while charging.

Arm payload is 2 to 3 kg. That is enough for light assembly, sorting, inspection and structured logistics tasks. It is not enough for the tote-moving and case-picking work that dominates warehouse humanoid pitches. Check the mass of the thing you actually want lifted before anything else.

Neither limit is a defect. They are what a machine of this size and price costs in physics. The failure mode is buying against a demo that quietly respected both.

The software stack

On EDU units the tooling is more serious than the price suggests. Unitree ships SDK2 in C++ with a Python interface, ROS 2 support over Cyclone DDS, official robot description files, a MuJoCo simulation model and an Isaac Lab reinforcement-learning repository. For a humanoid at this price point that is a real research stack, and it is the main reason the G1 has become the default platform in academic robot-learning labs.

The rough edge that shows up in practice is arm and locomotion coordination: the default stack handles walking well and handles manipulation well, and the seam between them, where the robot must move its body to serve its hands, is where people end up writing their own controllers.

For background on the pieces, see ROS 2, robot simulation and digital twins, and reinforcement learning for robotics.

The security problem nobody demos

This is the part that belongs in a purchasing decision and almost never appears in a review.

In September 2025, researchers Andreas Makris and Kevin Finisterre disclosed UniPwn, a vulnerability in the Bluetooth Low Energy Wi-Fi configuration interface used across Unitree's product line. Hardcoded cryptographic keys plus a trivial authentication bypass allow anyone within wireless range to obtain root-level control of the robot.

Three things make it worse than a typical embedded CVE:

1. It is wormable. A compromised robot can infect other Unitree robots within Bluetooth range without any user action, which is a robot botnet rather than a single-unit compromise.
2. The blast radius is the whole product line. Go2 and B2 quadrupeds, and G1 and H1 humanoids.
3. It was not patched. Responsible disclosure began in May 2025 and went public on 20 September 2025. As of May 2026, independent researchers reported no patch and the same authentication architecture still present in production firmware.

The practical mitigations are network-level, not vendor-level: keep units on an isolated VLAN, disable or physically shield BLE where you can, never operate them on a network you care about, and treat any Unitree robot in a shared space as an untrusted device. Our robot cybersecurity guide covers the general posture.

None of this means do not buy one. It means do not put one on your corporate network, and do not deploy a fleet in a public venue without understanding what an attacker in the lobby can do.

How it compares

The honest framing: the G1 competes on access, not capability. It is the humanoid you can actually get a purchase order approved for, and that is a real and underrated advantage when the alternative is a pilot contract with a two-quarter lead time.

Who should buy which tier

Buy Basic if you want a physical demonstration platform: a robotics course, a marketing booth, a museum exhibit, an executive briefing centre. You are buying a machine that walks and gestures on command. Accept that you will never write code for it.

Buy EDU if anyone in the building will write a line of software for it. That is every research lab, every robot-learning group, every startup prototyping a manipulation policy. The price jump is severe and it is the only tier that does what these buyers assume the robot does.

Buy neither if your task involves lifting more than 3 kg, running longer than two hours unattended, or operating on a network you cannot isolate.

Verdict

The G1 is the most important humanoid of this cycle, not because it is the most capable but because it is the first one a normal budget can reach. On the EDU tier it is a genuinely good research platform with real tooling, and the volume of published work built on it is the proof.

The reasons to go in with clear eyes are all knowable before you buy: the advertised price buys a robot you cannot program, the tier boundary is physical and permanent, the runtime is about two hours, the payload is a few kilograms, and there is a public, unpatched, wormable root exploit affecting every unit in the field. None of that is disqualifying for a lab. All of it is disqualifying for an unattended commercial deployment in 2026.

Frequently asked questions

How much does a Unitree G1 actually cost? In August 2026 the base unit sat in the low-to-mid teens of thousands of US dollars direct from Unitree, and 20% to 60% above that through resellers once shipping and import land. EDU research configurations are quote-only and ran roughly three to five times the base price. Unitree reprices often, so confirm before budgeting. What has not moved is the ratio: the developer tier is a multiple of the headline tier, not a markup on it.

Can you program the entry-level Unitree G1? No. The base model does not support secondary development. SDK access requires an EDU configuration, which is the single most expensive misunderstanding in G1 purchasing.

Can I upgrade a Basic G1 to EDU later? No. Motor mounts, wiring harnesses and compute module slots differ between tiers, so an upgrade means buying a second robot. Choose the tier on the first order.

What is the real battery life? About two hours of active use. Longer figures describe standby rather than walking and manipulation.

How much can the G1 lift? Roughly 2 to 3 kg of arm payload, which suits light assembly, sorting and inspection rather than materials handling.

Is the Unitree G1 safe to put on my network? Not without isolation. The UniPwn vulnerability disclosed on 20 September 2025 allows root-level takeover over Bluetooth from within wireless range, is wormable between nearby Unitree robots, and had not been patched as of May 2026. Use an isolated VLAN and treat the robot as an untrusted device.

What software does the G1 run? On EDU units, SDK2 in C++ with a Python interface, ROS 2 over Cyclone DDS, official robot descriptions, a MuJoCo model and an Isaac Lab reinforcement-learning repository.