The 20,000th embodied intelligent robot from AgiBot, the Expedition A3 Ultra, was officially delivered on September 24 at Hengqin Chimelong Spaceship Park.

On the same day, more than 300 AgiBot robots began regular duty across the park, stationed at over 100 interactive points and covering scenarios including performances, science education, guided tours, retail, hotels, and sports competitions.

From the signing of the agreement on August 23 to the official opening of the park, the two sides had only one month.

Wang Chuang, AgiBot partner, senior vice president, and president of the general business division, said during a media group interview that afternoon that the 300 robots correspond to over a hundred locations, with functions including narration, table tennis, basketball, golf, calligraphy, and rock-paper-scissors, and that the team also had to complete model upgrades, IP styling, material production, and on-site installation. "The biggest challenge was time."

The rollout of the 20,000th robot solved the problem of scale production. Once the robots entered Chimelong, they faced a different set of questions: whether they could operate stably every day, whether visitors would be willing to use them, and whether problems exposed on site could feed into the next round of model and product updates.

Scenarios are beginning to participate in "manufacturing" robots.

Shifting from delivery to operations

This batch of robots at Chimelong was acquired through a purchase model, with the two sides signing a 10-year strategic cooperation agreement and establishing a joint research institute for embodied intelligent cultural tourism.

Xiong Yan, AgiBot partner and co-president, broke down the "deployment state" into three requirements in the interview: operable, routine, and replicable.

Operable means that after deployment, the robot can continue to improve its capabilities; routine requires that the robot can go on duty every day without relying on one-off performances; and replicable means that the same capability can be copied from one robot to multiple robots, and then from one position and one park to more positions and parks.

Hengqin Chimelong Spaceship Park has 400,000 square meters of indoor space. Every day, the robots must contend with dense crowds, complex lighting, noisy sounds, and children, elderly visitors, foreign tourists, and dialect speakers.

Jiang Minling, vice president of Chimelong Group and general manager of Hengqin Chimelong Resort, explained that choosing an indoor park as the first large-scale scenario also took into account the current boundaries of the technology. Robots still have insufficient outdoor adaptability, and rainy-day applications require further research and development.

The real environment has already pushed AgiBot to adjust its products.

Wang Chuang introduced that the Expedition series has a battery life of 10 hours, compared with the industry's common level of about 3 hours. In the park scenario, a 10-hour battery life can cover most of the operating hours, usually requiring only one battery swap.

To cope with noisy environments, AgiBot optimized the acoustic front end, algorithmic noise reduction, multi-person recognition, and lip-shape recognition; the robots also support both Wi-Fi and dual 5G connections, and the park has allocated dedicated network slices for the robots.

Crowd density also raised safety requirements. The Expedition A3 added 360-degree perception; when the robot swings its arm, if someone suddenly enters the motion range, the arm will stop; compliant joint control is used to reduce force during accidental collisions. Wang Chuang also said that when the project was first deployed, staff would still be arranged for supervision, and improvements would continue based on operational data.

Visitors begin to change the product

The high-difficulty capabilities envisioned by researchers do not necessarily correspond to the projects visitors enjoy most.

Wang Chuang observed on site that the activities visitors lined up to experience might be robot rock-paper-scissors, tic-tac-toe, and table tennis. These projects have simple rules and direct interaction, allowing young children to participate and whole families to play together.

"Some things are not necessarily more liked by visitors just because they are technically difficult from a research and development perspective," he said.

In the next stage, AgiBot and Chimelong will continue to observe the usage rate and visitor feedback at each location. Projects with lower visitor interest will be replaced, while newly developed movements, models, and interactive content will be redeployed to the park.

Xiong Yan also mentioned that the park will assess the number of visitors experiencing each robot. Once the number begins to decline, the related program must be adjusted or eliminated. After robots have a general-purpose body and software capabilities, the cost of changing movements and program content is also expected to be lower than rebuilding a fixed set of equipment.

This operational feedback will also feed into model updates.

Wang Chuang gave an example in the interview: if a robot communicates with 1,000 visitors in a day, it may find 100 pieces of failure information. After model iteration, the team will observe whether the success rate for related issues can improve.

Xiong Yan said that after about one month of project deployment, the robots' interaction capabilities and adaptability to narration and guided-tour environments have been continuously changing, relying on operational data, model adaptation, and capability updates.

A recent research report by Nomura estimated that demand for high-quality physical interaction data in 2026 will be about 10 million hours, while the existing high-quality stock is about 500,000 hours. The report also mentioned that failure-recovery data is highly valuable, and most companies have still not closed the feedback loop after deployment.

A recent Citi survey of Mech-Mind provided another case: Mech-Mind has deployed more than 30,000 sets of 3D vision systems, focusing on collecting "near-miss" data during tasks, meaning samples that nearly failed, required intervention, or deviated from expectations, to continue optimizing models.

Whether and in what form data generated by the Chimelong scenario will enter general models was not disclosed in detail by AgiBot during this interview. What is already clear is that location usage rates, interaction failures, safety interventions, and visitor feedback are entering subsequent research and development and content adjustments.

Commercialization depends on the operations ledger

If robots can continue to update, operating costs must also be calculated.

At present, the Chimelong site has roughly 3 robots per staff member. Xiong Yan expects this state may last for several months, then gradually shift to 10 robots per person, with the ideal target being 20 robots per person.

Whether the human-to-robot ratio can decline depends on the robots' ability to autonomously go on and off duty, conduct group scheduling, operate safely, and support remote operations and maintenance. AgiBot and Chimelong are jointly deploying operations and maintenance systems, hoping to reduce the number of on-site supervisors.

Wang Chuang also listed other costs customers need to bear: purchase price, service life, network fees, token consumption, spare parts wear, and daily maintenance. Whether the equipment is used for two years or three years will directly affect the amortized result of equipment costs.

The revenue side also needs to be verified by operational data. AgiBot hopes robots can improve the visitor experience and increase repeat visits and in-park spending; Chimelong places more emphasis on the continuous renewal of content and performance experiences.

Wang Chuang said AgiBot will pay attention to how much cost the robots add to the park and how much revenue they bring. "Only when this thing gets better and better will our scale continue to expand."

Xiong Yan judged that it may still take 3 to 5 years for robots to reach a state where they require no debugging at all and can be moved between scenarios like a person. This does not prevent robots from first entering specific industries and continuing to improve efficiency, experience, and task capabilities through daily use.

The first production line determines how many robots can be produced. This "production line" at Chimelong, jointly formed by real visitors, operations staff, and the research and development team, will test whether robots can become more stable with use, whether supervision costs can continue to decline, and whether customers are willing to keep buying.