The archive · Space, Robots, Defence · Strategic decision · 2025–2026
TARS Raises Over RMB 5 Billion in 15 Months: Betting on End-to-End
Chen Yilun and team from Huawei, Baidu, DJI founded TARS in Feb 2025; angel $242M, Pre-A $455M, twice broke records.
TARS (它石智航)
What the business is
A company building embodied-intelligence robots: it developed the general-purpose embodied large model AWE3.0 in-house and uses robots and AI models to complete flexible manufacturing tasks such as wire harness assembly for factories.
Starting capital：Angel round of USD 242 million (Q2 2025) + Pre-A round of USD 455 million (April 2026), cumulative financing of over RMB 5 billion.
How it started
In February 2025, Chen Yilun (陈亦伦), former CTO and Chief Scientist of Huawei (华为)'s Car BU autonomous driving system, co-founded TARS (它石智航) with Li Zhenyu (李震宇), former head of Baidu (百度) Apollo/Apollo Go (萝卜快跑), Ding Wenchao (丁文超), a Huawei 'Genius Youth' (华为天才少年), and others. At Huawei, Chen had advanced end-to-end from a five-layer rule-based system to solving urban-village scenarios with fewer than 30,000 lines of code; in the second half of 2024, he concluded that the 'data + end-to-end' path could work for robots, so he formed a team to start the company.
What happened
In Q2 of the year the company was founded, it completed a USD 242 million angel round, setting the record for the largest angel round in China's embodied-intelligence sector; in April 2026, it completed a Pre-A round of more than USD 450 million, again setting the industry's highest single-round record. The team focuses on human-centered data collection (rejecting teleoperation data), released the world's first 'work-capable' general-purpose embodied large model AWE3.0, and uses wire harness robots and embroidery robots to validate generalization capability.
How it ended up
In 15 months since founding, cumulative financing exceeded RMB 5 billion; the wire harness robot completed a Guinness-record-level operation, customer expansion exceeded expectations, and the CEO said a 'small success' would be achieved in as fast as one year or as slow as two to three years.
Background
TARS (它石智航) is an embodied-intelligence startup founded in February 2025. Founder and CEO Chen Yilun (陈亦伦) is former CTO and Chief Scientist of Huawei (华为)'s Car BU autonomous driving system; co-founders include Li Zhenyu (李震宇), former head of Baidu (百度) Apollo Go (萝卜快跑), and Ding Wenchao (丁文超), a Huawei 'Genius Youth' (华为天才少年), among others. In 15 months, the company has raised over RMB 5 billion cumulatively, including a USD 242 million angel round and a USD 455 million Pre-A round, twice breaking China's all-time embodied-intelligence financing records.
The company's bet is: embodied intelligence must have a cross-robot, generalizable foundation model, and it can only be made to work through end-to-end learning plus large amounts of high-quality data. Chen Yilun (陈亦伦) validated this path at Huawei (华为)—solving urban-village autonomous driving with an end-to-end system of fewer than 30,000 lines of code, which traditional methods could not solve even with 300,000 lines of code; he judges that robots need at least 10 million hours of data, so he rejects teleoperation collection and chooses a human-centered 'imperceptible' data paradigm.
TARS's (它石智航's) first entry point is wire harness assembly on automotive production lines, which the team calls the 'fourth major line of industrial pre-assembly': the hardest flexible manipulation scenario across the industry, supported by about 1 million wire harness workers, and once solved it can generalize to lower-difficulty tasks. When it completed the Pre-A round in April 2026, the company had already released the general-purpose embodied large model AWE3.0 and shown deployment progress for wire harness robots and embroidery robots.
What has to be true
- The founding team brought conclusions validated for end-to-end in autonomous driving into robotics: the method works for autonomous driving and is likely to work for embodied intelligence.
- Data is the biggest bottleneck on this path; they bet that teleoperation data cannot scale and that a new human-centered data-collection paradigm is necessary.
- Wire harness assembly is the hardest and most labor-intensive single-point scenario in industry; solving it is equivalent to securing the high ground for a series of subsequent problems.
- Investors were willing to fund three consecutive rounds before the technology had a commercial closed loop; their bet was on the team and a high-density talent organization, not on existing revenue.
What can be applied
In unconverged technical routes, hardest scenario is clearest entry: prove end-to-end solves wire harnesses, then generalize to easier tasks.
Aftermath
As of April 16, 2026, TARS completed Pre-A over $450M, cumulative financing over RMB 5B, in technical breakthroughs and customer expansion. Wire harness robot can do manual labor; embroidery robot shows generalization. CEO Chen Yilun said 'small success' in 1-3 years; revenue and valuation undisclosed.
Sources
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card