Robotics in Factory Deployment: From Valuation Bubble to Line Integration

The number 2000 billion started to stand out when the market suddenly began to take the question of

1. The fall of Yu Shi has directly torn apart three 'Emperor's New Clothes'

1.1 Show-type valuation: A flip is not as good as tightening one screw more

The motion control capabilities of Yu Shi are indeed top-tier in the field of legged robots, from the consumer-level four-legged Go2 to the humanoid robot G1, the motion performance in public demonstrations has always been impressive. However, the capital market suddenly switched from

Do a rough calculation: even if the BOM cost of a humanoid robot is brought down to several ten thousand yuan, adding R&D amortization, channel fees, and after-sales costs, the selling price is hard to bring down much. But can it replace an ordinary worker in a factory? In the short term, no, because the cost is much higher than that of a worker. Can it replace an operator in a high-risk position? Yes, but the market for such positions is itself very small. This is the crack in showcase valuation: technological value is easy to appreciate, but hard to price.

I understand that many robotics companies in the past were pushed to do demos, compete in events, and hold press conferences because of the need for fundraising. But after the market sentiment reversed, this model no longer works—investors no longer pay for

1.2 Conceptual Valuation: Doing everything in every scenario is equivalent to doing nothing in any scenario

A few years ago, many robot companies' business plans would include a row of scenarios: home services, elderly care companionship, patrol inspections, commercial guidance, education and research, industrial handling... Each scenario sounded promising with demand, but when it came down to revenue structure, none of them could support a mature market capable of justifying high valuations.

After this round of decline, people began to re-examine the 'scenario ticket logic'. Some scenarios haven't even found their ticket: home robots, consumers are willing to spend how much money to buy a 'walking speaker'; elderly care companionship, who is the payer, the elderly, their children, or insurance institutions; education and research, the market really exists but is not sexy, can't sustain imagination. Conversely, the seemingly less sexy industrial scenarios are willing to pay real money for 'saving one person'. So you see the phenomenon that companies originally telling home stories and inspection stories are one by one heading into factories.

1.3 Hardware Millionaire-type Valuation: The whole machine looks good, but the core supply chain is still a thin bottom

There is also another piece of clothing that has been removed, which is the control over core components by many original equipment manufacturers. Yushu indeed has its own foundation in motors, joint modules, and motion control algorithms, but the maturity of the domestic robot supply chain has not yet reached a level where it can be trusted to operate without safeguards. When you take apart a legged or humanoid robot, harmonic gear reducers, planetary roller screws, six-axis force sensors, and high-power-density motors—each of these components, if blocked, will delay the delivery of the entire machine.

After this round of market sentiment cooling down, the most direct response from original equipment manufacturers is to 'solidify the supply chain.' They will develop their own technology if possible, and if not, they will deeply bind with suppliers. Because customers will not forgive delivery delays just because your appearance is cool. A day of production line stoppage is a day of lost money. Robot companies that only do the 'assembling' of shells and software will find it hard to tell a valuation story.

2. Why Saying 'Working in a Factory' is the Best Cool-Down Medicine for Robots

2.1 Factories Are the Most Structured Paid Scenarios

From industry trends, the first place where robots earn money is often not the most dazzling scenario, but rather the environment with the most constraints and the highest degree of structuring. Factories naturally have the advantage of 'information controllability': flat ground, controllable lighting, standardized production processes, and workers' actions with SOPs to learn. For robots, this is the best breeding ground to demonstrate the advantages of 'repetitive labor without fatigue and stable precision without fluctuation.'

Here, it's important to clarify one point: working in a factory doesn't necessarily mean that humanoid robots have to go in. The truly practical approach is to let more mature forms be implemented first: industrial robotic arms for loading/unloading, welding, and polishing, mobile robots for internal transportation, and composite robots that combine 'robotic arms plus chassis plus vision' into flexible workstations. Companies like Yushu, which specialize in legged robots, can enter the industrial sector by using legged platforms to replace traditional AGVs in complex terrains, such as cross-floor transportation, stairs, and inspections on uneven ground. This direction is more practical than forcing humanoid robots and can form orders more quickly in the short term.

2.2 Why 'Working in a Factory' Was Previously Unpopular, But Now Has Become the Standard

To be honest, 'working in a factory' was something many robot companies were not willing to openly talk about in the past. When the financing environment was good, everyone wanted to be a platform company, as if binding with a factory would mean becoming a laborer, with low gross margins, long payment terms, and fragmented demand. Factory clients are indeed difficult to please: today adding a material frame, tomorrow changing a process step, and the day after asking to increase the cycle time from 60 seconds to 45 seconds, and each manufacturer's process is different.

But now things have changed. First, the financing tide has receded, and companies must rely on revenue. Second, after going through a round of demo waves, everyone has realized that the final destination of all flashy scenarios is likely to be 'helping people work and saving costs,' which is exactly the most basic payment logic of factories. Taking the production line as a test bed is actually an advantage. The research and development direction has finally shifted from 'how to make robots look more like humans' to 'how to break down actions into cycles that factories can accept.'

2.3 The Mindset Shift from 'Tech Star' to 'Equipment Supplier'

Working in a factory, the corporate culture of a robot company also undergoes a transformation. Previously, gatherings were about the applause at product launches, now they're about being scolded by customers at delivery sites. But becoming a reliable equipment supplier is precisely the foundation for long-term business: customers know you're dependable, so they'll place repeat orders, request customization, and refer you to peers.

Recently, I visited several robot startups undergoing transformation. The business leaders put it plainly: 'Previously, we told investors about our vision; now, we tell plant managers about ROI, about how many people we can save in a year and how long it will take to break even.' This is clarity. After the valuation bubble has receded, companies that are still thriving are likely those teams that accepted the factory logic early on—small orders are fine, as reputation will snowball.

3. The real barriers of production lines lie in navigation, vision, communication, and safety

3.1 Navigation Selection: Reflective Tape, QR Codes, or SLAM?

Let's first discuss the most fundamental and easiest-to-fall-into-pitfalls of navigation for mobile robots. Currently, the mainstream options on production lines are three, each with its own suitable scope.

• Magnetic strips, reflective tape, and QR code navigation: low cost, high precision, fast deployment, but the downside is fixed paths and difficulty in later modifications. Suitable for production lines with stable processes that won't change lines in the short term. QR code navigation can achieve a precision of ±10 mm, and debugging is straightforward.
• 2D laser SLAM: No need to put anything on the ground, it completes mapping and positioning based on environmental features, offering high flexibility, with precision usually around ±20 mm. However, it is greatly affected by environmental changes. If the location of material racks changes or the state of stacked goods changes, the map needs to be maintained.
• 3D visual navigation: Can adapt to complex terrains and dynamic environments, suitable for stacking areas and mixed human-vehicle areas, but requires high computational power, complex calibration, and is expensive.

There is no 'most advanced' answer to selection; there is only 'most suitable.' I've seen a case: the workshop is full of metal shelves, and the customer initially insisted on using 3D laser SLAM, but metal reflections caused a lot of flying points in the point cloud, leading to occasional drift in positioning. Later, they switched to a combination of QR codes and magnetic strips, and it ran steadily for half a year. This conclusion may not be suitable for all factories, but it illustrates a principle: selection should first be honest about the scene, then responsible for the budget.

3.2 Scheduling System: Single-machine intelligence is useless, multi-vehicle coordination is the big test

Production line scenarios are usually not for a single robot working in isolation, but for dozens of AGVs or AMRs running simultaneously. At this point, even if single-machine navigation is strong, without a good scheduling system, the site will still be a mess. The more practical approach in the industry is to adopt open interface protocols like VDA5050, decoupling the vehicle end, scheduling end, and upper-level systems.

VDA5050 essentially uses MQTT messages to allow each vehicle and the scheduling center to communicate, with the message body including vehicle status, points, speed, load, and task status. The advantage is that vehicles from different brands can be connected to the same scheduling platform without private protocol binding. The disadvantage is that on-site debugging is quite involved: map points must correspond one-to-one with real workstations, avoidance strategies must be verified through simulation, and the fields sent by the upper-level WCS for tasks must align, otherwise, vehicles may get stuck at intersections for a long time. Many project delivery delays are precisely due to this seemingly minor环节.

3.3 Visual Guidance: Calibration and Lighting are More Troublesome than Algorithms

Robots working in factories can't avoid visual guidance: positioning and grasping, unordered loading, assembly alignment, and quality inspection all require robots to 'see' first. The four most critical words here are 'hand-eye calibration,' which solves the conversion relationship between the camera coordinate system and the robot coordinate system. If calibration is inaccurate, even the best algorithm is useless.

Common practices are divided into two types: the eye is on the hand, with the camera installed at the end of the mechanical arm; the eye is outside the hand, with the camera fixedly installed above the workstation. The calibration process is usually to print a calibration board, let the mechanical arm carry the camera through multiple poses, or let the fixed camera capture multiple angles of the mechanical arm end, and through matrix solving to obtain the external parameters. In practical operations, there are many details that affect precision: first, the calibration board must be flat and not reflective; second, the sampling poses must cover different heights and angles of the working space; third, the bracket for the fixed camera must be stable, and if the bolt is loose by one millimeter, the calibration is all for nothing.

Lighting issues are equally critical. I've seen a production line where, after adding a welding workstation nearby, the arc light caused the visual recognition rate to drop from 99% to 80%. Finally, they added a light curtain, replaced the light source with a polarized one, and changed the camera exposure from 'automatic' to 'fixed parameters,' and finally stabilized it. These kinds of issues can never be tested in a simulation environment.

3.4 Communication with PLC and MES: Robots Must Speak the Language of the Production Line

Robots entering the factory to work essentially involve upgrading existing production lines, so they must speak the language that the production line can understand. The main player in traditional production lines is PLC, with communication protocols often being Modbus TCP, EtherNet/IP, PROFINET, OPC UA. If the robot team only understands their own internal communication protocol, they will be stuck on-site.

My suggestion is straightforward: every robot team working on industrial applications should have at least one technician who understands PLC and OT networks. The robot must be able to perform interlock signals with PLC, respond to safety signals, and report its status to MES through OPC UA or MQTT. Many projects ultimately fail not because the robot itself is faulty, but due to miscommunication, protocol conversion issues, and confusion in data interface definitions. Additionally, many devices are now moving towards ROS2, and DDS's real-time communication capabilities are indeed strong. However, there are still a large number of old devices in the production line that only recognize the PLC system. The most common task in projects is actually protocol conversion and bridging, rather than training new models.

3.5 Safety Strategy: The production line won't give robots a second chance

Safety is the red line for production line acceptance, and there should be no侥幸. Collaborative robots and mobile robots are not just about stopping immediately when they touch a person. A complete safety design must at least consider speed monitoring, torque limits, setting up safety zones, deceleration zones, stop zones, emergency stop circuit redundancy, and the hard connection between the safety PLC and the robot itself.

Mobile robots, in particular, need to pay special attention to the blind spots of the safety radar, especially on the sides and rear of the vehicle. If a person approaches quickly from the side and is not detected, an accident could happen, and the entire project might be put on hold. A safe approach is to add ultrasonic or 3D vision sensors for additional coverage around the vehicle, and to associate the running strategy with speed and personnel density: reduce speed during times with more people, implement people-vehicle separation, and manage areas in zones. Safety is not the module that costs the most, but it determines whether you can even get the factory qualification.

4. Real-life pitfalls during two factory visits for debugging: details not mentioned in the drawings

4.1 The "hidden pitfalls" of the ground and environment are more common than algorithm issues

The first time I actually deployed a mobile robot on-site, I naively thought that mapping was just a few commands. However, I ended up in a big trouble. The workshop floor was made of grid steel plates, and the laser radar scanned it, resulting in fluctuating reflection intensity data; there were also a few forklifts running around, and the map was immediately messed up by dynamic obstacles. Later, the solution was: perform mapping during停产 periods, and mark dynamic areas as "high uncertainty areas" to reduce their weight; during operation, use real-time dynamic perception for avoidance.

Additionally, many factory floors appear flat but actually have a slope, even as small as a couple of degrees, which can cause odometry drift in wheeled robots. Do not trust the factory-calibrated wheel diameter parameters; always perform on-site measurements and calibration. Place the robot on a clean straight path for several meters and compare the endpoint error to reverse-calculate the wheel diameter correction factor. This step is almost always required for every new project.

4.2 Metal Reflection, Paint Lines, and Electromagnetic Interference: The Three Big Enemies of Vision Systems

The second project involved vision-guided pick-and-place operations, and the on-site issues were one more challenging than the last. The production line floor had yellow safety channel lines painted on it, and the industrial camera mistakenly identified these bright yellow lines as part of the workpiece edges; there was a welding machine next to the robotic arm, and the low-frequency magnetic field interference caused occasional signal jumps in the pulse signals; there was also a stainless steel column directly in the camera's field of view, creating a high-gloss band that led to frequent failures in template matching.

Finally, three solutions were applied to address these issues: first, adjust the ROI area to let the algorithm ignore the safety lines; second, replace the signal cables with shielded twisted-pair cables and re-ground them to resolve the electromagnetic interference; finally, add a polarizing filter in front of the lens to suppress glare, adapting to metal reflection scenarios. The entire process took two full days. These types of issues can never be detected in a simulation environment because simulations cannot replicate lighting, reflections, electromagnetic interference, and oil stains. This is also why I insist that

4.3 节拍优化不是调参数，而是拆动作

Factory acceptance is most concerned with the cycle time. Once, a customer required that a robot complete a picking, moving, and assembly cycle within 35 seconds. Initially, the cycle time was 52 seconds, and all modules were working normally, but it was just slow. I did not blindly increase the motion speed, but instead recorded the entire process as a video and analyzed it frame by frame, discovering that the main waste occurred in three areas:

• The mechanical arm was taking a明显冤枉路 (obvious detour) each time it returned to the safety point before picking;
• The waiting time for the pneumatic gripper to reach the correct position was set too conservatively;
• Visual capturing and image processing are executed sequentially, wasting 1.5 seconds.

The subsequent optimization was simple: change the return path to a smooth bypass; reduce the gripper waiting time from 0.3 seconds to 0.15 seconds; let the robotic arm trigger capturing during motion, reducing waiting time. The final cycle time was reduced to 33 seconds, and the customer nodded immediately. This experience made me realize that the core of cycle time optimization is motion analysis and process parallelization, rather than just increasing acceleration parameters. Increasing acceleration parameters can easily cause vibrations and overshoot, which is counterproductive.

4.4 The "Ghost Bug" of Communication Disconnection and Data Misalignment

There was another frustrating pitfall: after running for several hours daily, the robot occasionally lost communication during the night, and the human had to manually restore it in the morning, after which everything returned to normal. After a week of troubleshooting, the root cause was finally identified: during the night shift, large equipment startup caused a brief drop in power voltage, leading to the restart of the industrial switch.

Factory environments cannot be understood with the mindset of an office. On-site networks must use industrial-grade switches, paired with UPS; the 24V power supply in the control cabinet should be redundant, and communication modules and PLCs should be powered separately. These types of issues cannot be identified by just "looking at code and parameters." You must be on-site, checking logs, waveforms, and power supply. Therefore, I strongly oppose pure remote delivery of robot projects — it's not just about writing a program and uploading it. The gap between the field and the laboratory is much larger than many people imagine.

5. After the Foam Subsides, What Can Robot Companies Rely On to Survive?

5.1 Build a Fortress with Industry Know-how, Not Just Hardware

As robot bodies become increasingly similar, what truly sets them apart is industry-specific know-how in a particular process scenario. For example, with a grinding robot, knowing the process parameters for different materials, different sandpaper grits, and different contact forces can help customers achieve more consistent results, leading them to choose you naturally.

One of my friends who integrates collaborative robots has cut out all "everything" projects in recent years and focuses solely on a single area: screw fastening, pressing, gluing, and inspection. Now he doesn't lack customers; customers actively seek him out because his assembly process database has formed a barrier. This approach is worth all robot companies learning: instead of building a large, all-encompassing platform, focus on becoming the "line master" for a specific process step.

5.2 Let the Robot Run Data, Turning Work into Assets

One major benefit of working on-site is the ability to accumulate real operational data. For example, the temperature at which the robot slips while climbing, the number of times a joint experiences high torque impact, or the frequency distribution of recognition failures at a workstation — these data are something that can never be obtained in a simulation environment. They can both feed back into product iteration and serve as the foundation for predictive maintenance and value-added services later on.

In the long run, the robot companies that survive will not be those that only sell hardware, but those that can form a closed loop around "robot + data + service." Hardware is a one-time revenue, while data and maintenance services are recurring revenue, with the latter having much higher gross margins. Starting early is equivalent to gaining an extra one to two years of data accumulation cycle compared to others.

5.3 Operational Models and Talent Pipelines Must Change with "Working on-site"

Working on-site will also change the organizational structure of robot companies. Previously, it was research-driven, with algorithm engineers finishing coding and considering the job done; now it is delivery-driven, with the proportion of application engineers, debugging engineers, and maintenance engineers increasing significantly. Personally, I believe that in the next two years, the industry will truly lack not the people who design robots, but the application-oriented talents who can debug, maintain, and understand processes.

For individual professionals, this is also a clear opportunity. If you have some basic knowledge of robots and are willing to learn on-site processes, you will be more in demand than algorithm engineers who only know how to write code. This industry doesn't lack people who can "get the robot to run," but it lacks people who can "get the robot to work steadily and finish its shift."

After running a few rounds on the factory floor, my biggest takeaway is: robot industry valuations can recover, but trust needs to be built one robot at a time. Working on-site may not sound glamorous, but it transforms robot companies from "performers" into "service providers" — service providers may move more slowly, but every step is grounded. If your robot company is also going through a transition period, my advice is simple: hold fewer press conferences and go more to the factory floor; talk less about disruption and focus more on ROI; ask less about "how smart a robot can be" and more about "how much it can save the production line." When a company's engineers are willing to sit by the production line and watch the operation video frame by frame, that company is likely to survive this winter.