Want to Transform Factory Robots, First Be Transformed by the Factory

Text | Embodied Will, Author | Zhong Qingyuan, Editor | Yue Shisan

On September 30th, the American humanoid robot company Figure announced the final destination of Figure 02: a furnace in Finland.

The Figure 02 being processed is the company's previous generation product. With the expansion of the new generation product Figure 03 fleet, the company does not intend to continue maintaining Figure 02, and disassembling each unit individually would be too time-consuming, even slowing down the release of Figure 04. Therefore, most Figure 02 units have said goodbye, with only a few remaining in the headquarters.

However, before retiring, Figure 02 left some things behind.

According to Figure's disclosure, Figure 02 had operated over 1,250 hours in BMW's Spartanburg plant in the U.S., completing the loading of over 90,000 parts. The fault records left on the production line pointed to a weak spot: the forearm was the most frequently faulty part in the project.

Thus, for Figure 03, engineers redesigned the wrist electronics architecture, removing the original communication distribution board and dynamic cables, allowing the wrist motor controller to connect directly to the main computer. In June this year, Figure 03 has already returned to BMW's plant, demonstrating new logistics sorting tasks.

Figure 02 executing production tasks in the BMW plant. Image source: Figure

The old robot's body was melted down, but the issues it exposed in the factory have already entered the design of the next generation.

This also raises a question for robot companies: who decides what a good robot should be like?

Being able to move boxes does not equal being able to take over a loader position; standing on a production line does not equal being able to follow it continuously. When customers start paying for a specific job, the robot company's technical list needs to be reordered. Some capabilities must be added, some ideas need to be postponed, and even whether two legs are necessary has to be recalculated.

Humanoid robots have not yet transformed factories on a large scale, but factories have already begun rewriting the standards of a 'good robot.' Next, the real factor that may separate companies will be who can turn these on-site requirements into products faster and take responsibility for the work robots perform after delivery.

How to build a robot, the workstation decides.

The first round of transformation for robots in the factory is to require it to take over a job completely.

For customers, a capability only has a chance to become a reliable workforce if it truly covers the entire requirements of a job. If a robot can only handle part of a manual task, the remaining operations still need to be done by humans; if a robot has to stop charging after working for several hours, then production scheduling would have to adjust accordingly.

The Walker S1 from Ubtech was once limited by its operational range. In 2025, Jiao Jichao, the vice president and head of the research institute at Ubtech, admitted to the media that when developing this product, the company did not yet have a deep understanding of industrial scenarios, and the robot could only cover a partial area of the workstation, making it difficult to provide real application value to customers.

With the Walker S2, Ubtech added hip flexibility, allowing the robot to bend and squat to pick up objects from the ground.

There was also the issue of charging. According to Jiao Jichao's introduction at the time, the Walker S1 could work under high load for about three hours before needing an hour to charge. To address this downtime, Ubtech organized a dedicated team to develop autonomous battery-swapping technology, enabling the Walker S2 to swap its battery by itself.

Walker S2 reduces downtime during continuous operation through autonomous battery swapping. Image source: Ubtech

Being able to bend its waist and swap batteries on its own may not seem like reinventing the robot. However, for a specific workstation, these modifications determine how much work the robot can handle and how much work remains for humans.

Factory validation not only changed how robots should be designed but also changed what types of jobs Ubtech should prioritize first. Jiao Jichao introduced that the company had previously tried various workstations such as assembly, gluing, and so on. After validation, they gradually shifted their focus to handling, sorting, and quality inspection.

Choosing workstations is also choosing the order of research and development. Which tasks to do first and which capabilities to prioritize first now depend on what can truly be implemented on the ground.

In the BMW sheet metal loading project, Figure faced requirements specific to seconds: a task cycle of 84 seconds, with loading taking 37 seconds; the target correct placement rate per shift was over 99%, and the goal for human intervention was zero.

Such requirements would redefine what it means to 'be able to do it.' Once a robot is on the production line, putting parts in the designated location is not enough; it also needs to do it repeatedly and correctly within the specified time frame.

Going further, even the basic form of a robot can be influenced by factory requirements.

Wang He, founder of Galaxy General, said in an interview with the media in 2025 that the company had laid out a dual-leg design in its R&D, but at the time, factory and retail customers needed mobility, grasping, and placing capabilities, so the product chose a wheeled chassis. His reasoning was that legs would bring noise and range issues, while wheeled products could charge once every six to eight hours.

In June of this year, CATL disclosed that Galaxy General's heavy-duty robot Galbot S1 has entered its smart production line, handling operations in the production of modules and battery packs. When introducing this machine, the customer highlighted its 50 kg load capacity for its arms and an 8-hour battery life.

When a job can be completed with wheels, legs need to prove their additional value. Therefore, for robot companies, this is not just a question of choosing a chassis, but also a decision on which technology to introduce into products first and to face customers with.

From covering the entire workstation, to reducing charging downtime, to choosing a chassis, factories are rewriting 'what technology can do' into 'what this job really needs'. Companies that can make this transition have the opportunity to turn a single demonstration into a continuous business.

Delivery is just the beginning.

Robots have delivered the goods, but the company cannot finish its job. Customers want the work to be done every day, and this responsibility does not end with the robot's factory release.

Therefore, for robot companies preparing to enter factories, if they want to sell robots as labor, they must take on the responsibility of ensuring that labor can work continuously.

Agility, an American robot company, first showcased the potential of Digit in following autonomous vehicles to deliver packages.

In 2019, Agility announced a partnership with Ford: the car handles the transportation on the road, while Digit, the two-legged robot, takes over the last leg from the car to the delivery location. The car arrives at the door, the robot gets off and delivers the package to the doorstep.

A few years later, Digit, which had entered a multi-year commercial contract, stood in a warehouse.

In a warehouse operated by logistics company GXO in Georgia, USA, serving the clothing brand SPANX, autonomous mobile robots transport the boxes containing products, Digit takes over the boxes, places them on the conveyor belt, and sends them to the packaging workstation ahead. After a concept validation in 2023, GXO signed a multi-year agreement with Agility in 2024 to deploy Digit and the accompanying software Arc as a robotic service.

Digit is operating in the GXO warehouse on a daily basis. Photo source: Agility Robotics

Receiving the box, turning around, placing it down, and then receiving the next box. This repetitive action still needs to eliminate unnecessary steps.

When introducing improvements after one year of deployment in GXO, Agility disclosed that the team redesigned the navigation system, allowing the robot to find the path with the fewest steps, reducing unnecessary actions during starting and sharp turns, and improving the efficiency of moving with a load. It needs to carry a load, repeatedly stopping and turning around in narrow spaces.

The production line is not always smooth. If the downstream workstation is backed up, or employees are on break, Digit will first place the boxes aside, allowing the delivery mobile robot to continue passing through, and then replenish the stored boxes once the conveyor belt resumes.

At this point, Digit's task is no longer just moving boxes from one place to another, but rather a position within the production process. It needs to judge based on the status of upstream and downstream processes, deciding when to continue and when to wait.

On September 15, 2024, Agility released Digit 5. The company directly attributed this revision to the requirements received after the previous generation had been working in customer production sites for nearly three years.

According to the specifications published by the company, to take on more carrying tasks, Digit 5 increases its load capacity and reach height, and improves the ratio of work time to charging time from the previous 2:1 to 10:1, achieving '9 minutes of charging, 90 minutes of work'. At the same time, the new product also improves personnel detection and new safety control designs, aiming to reduce the reliance on isolation facilities during deployment.

However, improving the body only completes part of the delivery.

Agility's contract with GXO from the beginning included the accompanying software Arc. This platform handles mapping, workflow configuration, operation management, and troubleshooting, integrating Digit into existing automation equipment. With Digit 5, Agility further emphasized that through Arc, customers can view working time, throughput, and fault intervals, and connect warehouse management, warehouse execution, and manufacturing execution systems.

All these tasks require Agility to continuously invest human resources. Maps need to be built, tasks need to be configured, software needs to be connected, and faults need to be handled. On the surface, customers purchase a machine, but in reality, Agility takes on a continuous responsibility.

This continuous service has also entered Agility's pricing model. Customers can pay for deployment fees and ongoing robot service subscription fees, or they can purchase Digit and then buy software and maintenance services annually.

From the initial vision of delivering packages with Ford to the current task of moving boxes in GXO's warehouse, Digit has found a specific job. Around this job, Agility has expanded what it delivers to include enterprise software, system integration, and ongoing services.

It needs to be sold and stay long-term.

If robots continue to work in real environments for several years, the problems continuously exposed by customers will eventually change the product into what shape?

Quadruped robots, which entered inspection sites earlier than humanoid robots, may already provide a reference.

In late 2019, Yunchudeshen collaborated with China Southern Power Grid's Digital Grid Research Institute to have the first generation of Jieying X10 autonomously inspect substations. A requirement that emerged later was that it had to work even on rainy days, which led to the X20, released in 2021, adding waterproof and dust-proof capabilities, moving closer to all-weather inspection.

The X30 later faced even more detailed issues. Yunchudeshen mentioned in the product description that dust, dirt, and changing light conditions affect the robot's autonomous charging, so they redesigned the charging positioning solution; facing customer demands for long-term operations, it corresponds to longer battery life and quickly replaceable batteries.

These changes ultimately face the test of robots staying in real-world environments for the long term. The Ningxia unmanned wind farm project disclosed by Yunchudeshi in 2025 is about 10 kilometers away from the nearest populated area, making it inconvenient for people to arrive on site at any time, so the stability of robots in completing inspections becomes even more important.

X30 performs unmanned inspections in the Ningxia Gobi unmanned wind farm. Image source: Yunchudeshi

In this project, X30 can work with drones to achieve two inspections per day, each lasting about 80 minutes, checking equipment such as transformers and isolating switches, and transmitting the results back to the platform.

The challenges of quadruped inspection and humanoid carrying are different, but rain, dust, and charging docking have one common point: they are rarely the most shining capability of a robot, yet they may decide whether customers dare to continue assigning tasks to it tomorrow.

The requirements for humanoid robots in factories are also becoming more specific in this direction. Sales and revenue can indicate that the market is willing to pay, but what these robots are ultimately used for and whether they can complete a job long-term still need to be tested by the production site.

YuShu has turned humanoid robots into a business with scale. Its August 2026 IPO filing disclosed that the revenue from humanoid robots in 2025 was about 868 million yuan, accounting for 51.78% of main business revenue, surpassing quadruped robots.

According to the company's previous inquiry reply disclosure of the total data by field for the four quarters of 2025, scientific research and education account for about 76% of the revenue from humanoid robots. Here, scientific research and education also include technology companies and developers purchasing robots for secondary development, research, and model training. A considerable portion of these demands involves purchasing platforms for continuing the development and research of new capabilities.

However, when robots truly enter production lines, what customers want to buy is a job that needs to be completed on time. Both types of demand can support business, but being good at selling robots to developers does not automatically win the recognition of production customers.

Selling a robot may mean that a research and development project has just begun; having a robot stay on the production line means it has to repeatedly deliver work results. The industry cannot use the prosperity of the former market to prematurely declare the victory of the latter market.

Looking further ahead, the capabilities worth betting on in the next round of competition may not all be hidden in more eye-catching actions. Who can turn faults into the design of the next generation, and turn a customer's process into a product that can be continuously delivered, will have a better chance of retaining the first-mover advantage.

Figure 02 has withdrawn, but the problems it left behind in the BMW production line have entered Figure 03.

Jumping into the furnace is enough to be exciting, but what's more worth continuing to watch is how production lines are being redesigned for robot companies.Return to the homepage, view more