Most humanoid (人形机器人) startups in China spent 2024 and 2025 arguing about how to make a machine walk like a person. Wang He (王鹤) asked a blunter question: does it need legs to start earning its keep? The answer shaped Galbot (银河通用), the Beijing embodied-AI company he founded in 2023 — and it made him, at 34, one of the most funded robotics builders in the country.
A professor who refused to wait for the lab
Wang He was born in 1992. He earned his bachelor's degree in electronic engineering from Tsinghua University in 2014 and a Stanford PhD in 2021, where he worked with Leonidas J. Guibas, a member of three U.S. national academies and a giant in 3D computer vision. He then returned to China to become an assistant professor at Peking University, where he founded and leads the Embodied Perception and Interaction Lab (EPIC Lab) and directs the embodied-intelligence center at the Beijing Academy of Artificial Intelligence (BAAI).
In May 2023 he started Galbot (银河通用). The academic pedigree is real, but the bet was practical: build robots that can perform useful, repetitive physical tasks in controlled environments now, and let the bipedal elegance arrive later.
A wheeled robot that chooses utility over looks
Galbot's first product, the Galbot G1, launched in June 2024, deliberately broke from the humanoid fashion. It rides a wheeled base with two arms and a body that folds and rises — about 173 cm tall, with an arm span near 190 cm and the ability to reach roughly 240 cm high and lift its torso about 65 cm. Those numbers are not about imitating a person; they are about reaching the top and bottom shelves in a store.
As tracked by BrainNet (脑机网), China's AI observatory, Galbot's design philosophy is the opposite of the "make it walk" crowd: a wheeled chassis is steadier, cheaper to run, and lets the team pour engineering into the actual bottleneck — a robot arm that can pick up objects it has never seen before.
The data paradox he claims to have solved
Robots that manipulate the real world face a brutal shortage of training data; you cannot scrape "how to stock a shelf" from the internet the way you scrape text. Wang He's answer was a training recipe Galbot repeats like a mantra: synthetic simulation data for pre-training, real data for post-training — simulation-first, real-world-second.
Out of that came a family of vision-language-action models:
- GraspVLA, an end-to-end embodied grasping foundation model.
- GroceryVLA, tuned for retail picking.
- TrackVLA, a product-grade navigation model.
The promise is generalization — a robot that can grab an unfamiliar item without being hand-taught object by object. That is the difference between a machine that needs a demo video for every task and one that can be dropped into a new store and get to work.
Funding at a record pace
Galbot's capital story is the part outsiders find hardest to believe. Reported rounds include:
- An angel round of about RMB 700 million (≈ US$97M / HK$760M) in June 2024, with backers including Meituan's investment arm, BAIC, SenseTime's fund, and iFlytek's fund, plus top venture firms.
- A strategic round of about RMB 500 million (≈ US$70M / HK$550M) in November 2024.
- An RMB 1.1 billion (≈ US$153M / HK$1.2B) round in June 2025 led by battery giant CATL (宁德时代) and Pujan Capital — reported at the time as the largest single financing in China's embodied-AI sector.
- A further RMB 2.5 billion (≈ US$348M / HK$2.7B) round in 2026 led by the National AI Industry Fund (国家人工智能产业投资基金), with China Petroleum, CITIC, Bank of China, and SAIC among the investors.
Across these, Galbot's cumulative financing reached into the tens of billions of RMB and its valuation surpassed RMB 20 billion (≈ US$2.8B), placing it among the highest-valued private humanoid (人形机器人) companies in China. Notably, CATL has deployed Galbot's S1 robot for autonomous work inside its own battery factories — a rare case of an investor using the product in a real production line.
From retail shelves to the Spring Festival Gala
Galbot first proved itself in unglamorous places: smart-retail stores in roughly ten Beijing locations, where robots handle stocking and picking, and a joint venture with Bosch to push into industrial settings. Then came the spotlight. For the 2026 CCTV Spring Festival Gala (春晚), Galbot was named the designated "embodied large-model robot," billed as the first robot on that stage that actually does tasks — sorting, fetching, and handling objects — rather than performing a choreographed dance.
What readers can do now
- Separate "can walk" from "can work." When evaluating robotics firms, ask which ones have paid deployments in retail, warehouses, or factories today. Galbot's wheeled, task-first path is a useful lens.
- Follow the synthetic-data claim. The "simulation-first" training recipe is the make-or-break idea; watch for peer-reviewed benchmarks or third-party tests of GraspVLA-style generalization, not just launch videos.
- Watch the CATL relationship. An investor running the robot in its own factories is the strongest real-world validation signal — track whether that expands beyond pilots.
Honest limitations
This profile relies on Chinese state media (People's Daily), Forbes China, The Paper, and trade coverage available through late 2026; it does not include Galbot's private financials or audited results, which are not public. Funding figures are reported round sizes and may be rounded, overlapping, or revised in later filings. I have not independently verified robot performance metrics, the exact size of Galbot's dataset, or the clinical/industrial efficacy of deployments — those require access I do not have. My framing treats Wang He's simulation-first approach as technically credible, but rivals pursuing bipedal, real-data-heavy, or teleoperation-assisted paths may prove more robust in unstructured environments. The "highest-valued" claim depends on which private firms are counted and on dates of measurement.