A PhD student in Boston unboxes a full-size humanoid for less than the price of a used car, plugs in a 3D LiDAR, and starts teaching it to walk. Five years ago that sentence was science fiction. Today it is a purchase order from Unitree (宇树科技), the Hangzhou company that has quietly become one of the highest-volume humanoid makers on the planet.
The G1 is not a household butler. It is something arguably more disruptive for the field: a cheap, capable, open-enough research platform that lets thousands of teams actually break things and learn.
Key takeaways
- The G1 base model starts at US$13,500 (EDU variants run higher), roughly 90% cheaper than Unitree's earlier H1 humanoid at about US$150,000.
- Standing 1.32 m and weighing about 35 kg, it offers 23–43 degrees of freedom depending on configuration.
- A 9,000 mAh quick-swap battery gives roughly 2 hours of runtime; perception comes from a 3D LiDAR plus a depth camera.
- Independent trackers place global shipments above 5,500 units by mid-2025, with deployments at Stanford, MIT, UT Austin and Amazon research.
How a quadruped company rewrote the price curve
Unitree made its name with four-legged robots — the Go1 and Go2 — before entering humanoids. That lineage matters. The G1's affordability comes from reusing proven actuators, motors, and manufacturing know-how developed for quadrupeds, then scaling volume. The company's own store lists the base G1 from US$13,500; the predecessor H1 launched near US$150,000. That gap, not any single spec, is the story.
At 1.32 m tall and 35 kg, the G1 is compact enough to fit through doorways and reach desk height, yet robust enough to walk at about 2 m/s, run faster, and — in demo footage — perform kip-ups and dynamic gestures. Degrees of freedom scale from 23 in the base version to 43 in the EDU flagship, with optional force-controlled three-fingered Dex3-1 hands.
Built to be taken apart
Where many humanoids are sealed showcases, the G1 is explicitly a developer object. The EDU version supports ROS 2, Python and C++, and can mount an NVIDIA Jetson Orin compute module. Unitree ships a software development kit and promotes its UnifoLM (Unified Robot Large Model) for imitation and reinforcement learning, with simulation-to-reality training pipelines.
The hardware reinforces the message: a quick-release 9,000 mAh battery, dual encoders on the joints, industrial-grade crossed-roller bearings, and a foldable frame (69 cm when folded) that makes shipping and storage trivial. For a lab ordering ten units, that fold-and-ship practicality is as valuable as any benchmark.
Where the robots actually are
Unitree's volume claim is the part worth scrutinizing. Independent robotics trackers put global G1 shipments above 5,500 units by mid-2025, and the company's machines appear in research settings from Stanford and MIT to UT Austin, with Amazon cited among early robotics-research users. These are not consumer homes; they are labs treating the G1 as a standardized testbed — exactly the adoption pattern that builds an ecosystem.
That ecosystem effect is the real moat. When thousands of teams run the same platform, they produce shared gait libraries, control forks, and failure reports. Unitree has even field-tested extremes — cold-environment walks in Xinjiang's Altay region — that double as credibility marketing for the actuator and battery design.
The honest caveat: it is not autonomous yet
Almost every impressive G1 clip involves a remote operator or a pre-scripted sequence. Full task autonomy, safe unsupervised operation around people, and plug-and-play manipulation remain unfinished. Unitree itself warns that the robot requires trained operators and is not certified for safe human-collaborative interaction like a true collaborative robot. At 35 kg moving at up to 7.2 km/h, an unsupervised G1 is a liability, not a helper.
There are also harder questions. Security researchers have flagged vulnerabilities in internet-connected robots, and the same low price that democratizes research also lowers the barrier for misuse. Unitree states its sales are civilian-focused, but a platform this accessible will test that boundary.
Why the cheap humanoid changes the field
For decades, humanoid research was the domain of a handful of well-funded labs because the hardware cost millions. The G1 collapses that gate. A university teaching lab, a startup, or an overseas hobbyist collective can now iterate on the same physical platform the frontier labs use. That compresses the time between an idea and a working prototype — and it is why "who ships the most" may matter more than "who demos the best."
What readers can do now
- If you lead a robotics lab, benchmark the G1 EDU against costlier platforms for locomotion and manipulation research; the savings can fund more experimental runs.
- Developers should treat the G1 as a ROS 2 / Python testbed and contribute to shared control repositories rather than rebuilding from scratch.
- Buyers should budget for safety infrastructure ( cages, e-stops, trained operators ) — the sticker price is only part of the real cost.
Honest limitations
Figures here come from Unitree's official product page (pricing, specs), independent robotics databases and tracker sites (shipment counts, deployment lists), and secondary reporting on cold-environment tests. The "5,500+ units shipped by mid-2025" number is from a third-party tracker, not an audited Unitree disclosure, so treat it as indicative rather than exact. I have not handled a G1 unit myself; performance claims about autonomous task success in clinical or needle-procedure trials attributed to the platform by some outlets were excluded because they were not verifiable against a primary specification. My stance is that of an external observer of China's robotics industry, not a Unitree affiliate.