Zhihuijun's robots start work
稚晖君的机器人，开始上班了
B · Useful, limited China Entrepreneur magazine interviews AgiBot's founders and executives on real deployments, with useful operating figures, but nearly all numbers are the company's own and unaudited.
Summary based on the full text·Original: about 10 minutes to read
In short
A 6 November 2026 feature from China Entrepreneur magazine (中国企业家), carried by 36Kr, on AgiBot (智元) after its 29 October launch event. Co-founder Peng Zhihui ("Zhihuijun") introduced "Worken", a unit for a robot's effective working hours, and said robots are moving from development to deployment. The article tests that claim: AgiBot has built 20,000 robots, but its own SVP expects only a few thousand real deployments this year. It covers a 3C factory line, a theme park and a metro system, and the cost of keeping engineers on site.
Key numbers
- over 60% vs under 20%share of global humanoid robots in H1 2026 going to labs, exhibitions and tourism versus industry (third-party count cited)
- 64,824 picks, 17,625 tabletsoutput of the AgiBot robot line at Longcheer during a six-day livestream, 22 s average takt
- 1:1 to 3:1robots per engineer at Chimelong park so far; target 10:1 by year end and 20:1 next year
- 20,000thAgiBot robot off the line in September 2026; an SVP says a few thousand real deployments this year would be very good
- 10% to 20%share of trajectories usable from blind data collection, versus 80% to 90% from in-scene tasks
As reported by the source.
Chinese·Web page·opens 36kr.com
What it reports
- Per a third-party count, over 60% of humanoid robots shipped worldwide in H1 2026 went to labs, research, exhibitions and tourism; under 20% to industry.
- Longcheer 3C line, in volume production since March: in a six-day livestream robots worked over 64 hours, made 64,824 picks and produced 17,625 tablets at a 22-second average takt, said to match people. The line allows at most one error per 10,000 operations.
- Counterpoint: AgiBot shipped 9,700 robots in H1 2026, over 43% share. H1 revenue close to RMB 2bn; Hong Kong listing started.
- Chimelong park (Zhuhai): over 300 robots at more than 100 points; at first one engineer per robot, which cost more than human staff. Ratio now 3:1, targets 10:1 and 20:1. Over a million visitor interactions in the first month.
- Guangzhou Metro runs over 120 G2 robots for security guidance, rule-breaking detection, Q&A and patrol.
- Semiconductor packaging work requires a 99.999% success rate.
- R&D staffing: 30% on hardware, 70% on intelligence. Training cluster near 10,000 cards this year.
- Data economics: a standard dataset of a million hours may be priced around RMB 200m in two to three years; blind collection yields 10-20% usable trajectories, in-scene collection 80-90%.
- Most customers do not yet use the 10-plus degree-of-freedom tactile hand in real work because success rates trail grippers. Executives cannot yet define the "value density" that Worken billing would need.
How they got it
Interviews with AgiBot executives at and after the launch, plus a cited Counterpoint figure and an unnamed third-party count. No customer was interviewed independently.
Limits
- Company figures throughout; the livestream output and deployment ratios are not independently checked.
- "At most one error per 10,000" is stated as the line's requirement, not a measured error rate.
- The third-party source for the 60% / 20% split is not named.
- Worken is a marketing unit with no published definition or pricing.
Why it matters for companies
The useful numbers are about operations, not robots: one engineer per robot at launch made the deployment more expensive than human staff, and scale depends on reaching 10 or 20 robots per support person. Service settings deploy faster than factories because customers accept value that is not a strict takt-time win. Factories that keep process data on site need training servers inside their walls. Anyone pricing humanoid pilots should ask for support ratios and real working hours, not units shipped.
Key passages in translation
Chinese original
最开始几乎是一个机器人配一个工程师，300个机器人，就得有300个工程师入场。这样一来，成本比原来纯人工还要高。
At first it was almost one engineer per robot; 300 robots meant 300 engineers on site. That made the cost higher than using only people.
Chinese original
今年真实部署能有几千台，就很厉害了。
If we get a few thousand real deployments this year, that will be very good.
Chinese original
他也坦承，现在说不太清楚的就是价值密度。
He also admitted that value density is what nobody can yet explain clearly.
Summary based on the full text. Read the original for the full argument and data.Read the original on 36kr.com