A theme park where robots are not a show but the staff: that is the experiment running at Chimelong Spaceship Park in Hengqin, Zhuhai, and its first month of data is now public.
More than 300 robots have been on normalized duty for over a month, and visitors have interacted with them more than a million times.
What the operator does with the underperforming ones is the most revealing part of the story.
A Park Full of Robots on Regular Shifts
Per a CCTV report on November 15, Chimelong Spaceship Park has deployed more than 300 robots across more than 100 interactive points, spanning seven scenario families: sports competition, intelligent companion touring, science education, guided tours and shopping assistance, performances, hotel service and robot competitions.
The machines are AgiBot (智元机器人) units across the company's lineup, humanoid (人形机器人), wheeled and quadruped, on duty since late September under a strategic cooperation signed in August 2026. The popular attractions are competitive and physical: gomoku and table tennis against a machine, pitch-pot toss games, robot football goalkeeping. At peak visitor flow, these stations draw queues.
The science education line has become an unexpected hit. Robots serve as teaching assistants in study courses for students, and at peak periods daily sign-ups exceed 4,000, according to park general manager Gao Rui.
The Numbers Behind One Month
AgiBot co-founder Xiong Yan gave CCTV the operating totals: by late October, cumulative robot-visitor interactions had surpassed one million. During the seven-day National Day holiday alone, the fleet ran 12,000 hours and completed 299,000 voice interactions.
Some back-of-envelope arithmetic on those figures, ours rather than the operator's: 12,000 hours across seven days and roughly 300 robots works out to about 5.7 hours of duty per robot per day. The 299,000 voice interactions spread the same way gives roughly 140 interactions per robot per day, or one exchange every two to three minutes of continuous high-traffic dialogue. That is a convenience-store intensity of conversation, not a showroom demo schedule.
The hardware numbers are consistent with that load. The main models run six to eight hours continuously, some exceed ten, and a dual-battery design supports fast swapping without taking a unit off duty. Robots enter their work zones in formation before opening, with operators confirming software and hardware status.
Running 300 Robots Like 300 Employees
The park has built what it calls the first one-stop embodied-intelligence operations and maintenance management platform, putting humanoid, wheeled and quadruped units on a single screen. Every unit is monitored for on-duty status, faults and daily interaction counts.
The discipline extends to performance review. If visitor engagement at a station declines, the robot there gets adjusted or removed. The park has collected more than 720 optimization suggestions and resolved over 640, according to CCTV.
This is the part worth pausing on. Most robot deployments are judged by whether they work at the ribbon-cutting. This one is judged by whether each unit earns its station, week after week, in a venue with daily crowds in the tens of thousands, children approaching from every direction, music and announcements in the background, and interruptions mid-dialogue. Xiong Yan's argument to CCTV was that one month of high-intensity real operation directly validates normalized deployment in high-traffic, high-complexity environments.
Gao Rui said the park's phase one was planned at under 300 units, but demand outran the plan, and phase two will scale to nearly a thousand robots. Scaling to a thousand is less a manufacturing question than an operations-radius question: battery swaps, program updates, fault handling and station content refreshes all need a bigger maintenance system, not just more units.
Why a Theme Park Is a Good Proving Ground
Theme parks sit in a useful middle ground for embodied intelligence (具身智能). They are more forgiving than a factory line, where a two-second hesitation stops production. They are harsher than a hotel lobby, with higher traffic, more unpredictable human behavior and longer duty cycles.
For the manufacturer, the park is a paid data pipeline: every failed interaction, battery swap and edge case feeds the next model iteration. For the operator, robots shift the cost structure of attractions from one-time rides to updatable software-backed stations. For the industry, the million-interaction figure is a first public benchmark for consumer-facing robot endurance outside the lab.
The honest benchmark to watch is not the phase-two headline. It is whether any single station proves valuable enough that the park adds units to it, the way a store restocks a product that sells.
Honest limitations
- Nearly all operating figures, including the one million interactions, 12,000 holiday hours and 299,000 voice interactions, come from a single CCTV report quoting the two companies; independent verification of interaction counting methodology is not available.
- The per-robot daily figures in this article are our own arithmetic from aggregate numbers and assume roughly 300 units throughout; actual fleet sizes may have varied.
- The phase-two scale-up to nearly 1,000 units is a stated plan, not a funded, scheduled commitment.
What readers can do now
- Watch for a published operations report after the next peak season; sustained per-robot uptime would be stronger evidence than launch numbers.
- If you evaluate robot vendors for high-traffic venues, ask for three figures: duty hours per unit per day, interactions per unit per day, and stations expanded versus retired after review.
- Treat normalized operation claims as verified only when an operator discloses utilization and maintenance data alongside them.
