Manufacturing Transformed by Physical AI: From "Thinking AI" to "Acting AI" [Part 8]: Use Case 1: Assembly, Picking, and Processing—AI Overcoming Irregular Shapes, High-Mix Production, and Setup Changes—
Up until the last installment, we looked at cases in the US and Europe, as well as patterns of success and failure. So, what kind of tasks is Physical AI specifically suited for?
Starting with this installment, we will look at "use cases" in detail over three episodes. Part 8 covers assembly, picking, and processing, which are the fundamental tasks of manufacturing.
How does Physical AI overcome the challenges that conventional robots struggled with, such as "irregularly shaped objects," "high-mix low-volume production," and "frequent setup changes"? We will explain this with real-world examples from companies like Foxconn, Fanuc, and Covariant.
1. The Revolution in Handling Irregularly Shaped Objects
1-1. Fusion Technology of Vision AI and Tactile Sensors
Conventional robots were good at handling "things with fixed shapes." Metal parts and plastic molded products have constant shapes, and their positions can be predicted.
However, "irregularly shaped objects" were difficult. Vegetables, clothing, and soft packaging materials all have different shapes one by one. Even for the same product, the appearance changes depending on how it is placed.
Physical AI solves this challenge by fusing "Vision AI" and "tactile sensors".
Vision AI scans the target object with a 3D camera. It instantly recognizes the shape, size, and orientation. What is important is that it can identify "things it has never seen before." Even for vegetables that have not been taught in advance, it understands that "this is a vegetable."
Tactile sensors are attached to the robot hand. They detect the force, slippage, and deformation when grasping an object. For example, when grabbing a tomato, if the force is too strong, it will be crushed. If it is too weak, it will drop. Tactile sensors automatically adjust the appropriate force.
What fuses these two technologies is modern Physical AI. It judges "what to grab and how" with vision, and adjusts the "actual gripping force" with touch. AI reproduces what humans do unconsciously.
The technical breakthrough came around 2020. Due to the evolution of deep learning, the accuracy of Vision AI improved dramatically. At the same time, small, high-performance tactile sensors were put into practical use. Combining these made the handling of irregularly shaped objects possible.
Figure 1: Fusion of Vision AI and Tactile Sensors
1-2. Foxconn (Taiwan): Application to Electronic Device Assembly
Foxconn (Hon Hai Precision Industry) is the world's largest EMS (Electronics Manufacturing Services) company. It manufactures iPhones, iPads, PlayStations, and electronic devices all over the world.
Foxconn has been promoting a "million-robot plan" since around 2015. Initially, it was a plan to introduce a large number of conventional industrial robots. However, the complexity of electronic device assembly became a barrier.
Electronic device assembly involves hundreds of parts. Screws, connectors, cables—the sizes and shapes vary. Furthermore, product models change frequently. Every time a new iPhone comes out, the assembly process changes.
With conventional robots, teaching was required every time the model changed. This work, taking several weeks, occurred several times a year. This does not increase efficiency.
Since 2020, Foxconn has accelerated the introduction of Physical AI. In its factory in Shenzhen, China, robots equipped with AI vision systems are in operation.
These robots identify parts by "looking" at them. They recognize screw types, connector orientations, and cable routing positions, all with cameras. Even when new parts arrive, the AI learns automatically. Teaching time has been reduced from several weeks to several hours.
What is particularly innovative is the "cable routing" task. Routing flexible cables to precise positions on a circuit board was impossible with conventional robots. Cables are soft and their shape is not fixed.
Foxconn's Physical AI robots sense the "flex" of the cable with tactile sensors. They route the cable with appropriate tension, neither pulling too hard nor leaving it too loose. The success rate is about 95%.
Foxconn develops this technology in-house. Instead of relying on external vendors, it has an AI research team within the company. This allows it to integrate manufacturing know-how with AI technology.
Chart 1: Introduction Effects at Foxconn
1-3. Handling Flexible Objects in the Food and Apparel Industries
Handling irregularly shaped objects is most effective in the food and apparel industries.
In the food industry, there is a case of a major US chicken processing company (anonymous). This company introduced Physical AI for chicken cutting tasks.
Chicken is a typical example of an irregularly shaped object with large individual differences. The size, shape, and fat distribution are all different. Skilled workers judge the optimal cutting position by looking and touching.
The AI vision system fully grasps the shape of the chicken with 3D scanning. It analyzes the position of bones, the run of muscles, and the distribution of fat, all in a few seconds. Then, it calculates the optimal cutting position.
The robot arm is equipped with a sharp knife.With tactile sensors, it senses the resistance when cutting. If it hits a bone, it fine-tunes the angle. As a result, the yield has improved.
The introduction effect was clear. Processing speed improved by 20%, and yield improved by 3%. This is an increase in profit of several hundred million yen per year. Furthermore, the working environment was improved. Workers were freed from heavy labor in refrigerated environments.
In the apparel industry, the logistics center of a major Japanese apparel company (anonymous) is a case study. This company introduced Physical AI for the inspection and sorting of returned clothing.
Returned clothing is extremely diverse. T-shirts, jackets, skirts—the materials, colors, and shapes are all different. The folding methods are also inconsistent. These need to be sorted by size and color.
AI vision "reads" the clothing. It recognizes the tag and acquires product information. At the same time, it also performs visual inspection. It automatically detects stains, tears, and missing buttons.
The robot hand has flexible "fingers." It gently grabs the clothing, spreads it out to check, folds it, and sorts it. The success rate is about 90%. If it fails, a human handles it.
At this company, 70% of the sorting work was automated. Workers can focus on the remaining 30% of difficult cases and final confirmation. Overtime during busy seasons has been significantly reduced.
Chart 2: Examples of Irregular Object Handling by Industry
2. Realization of High-Mix Assembly Lines
2-1. Responding to Lot 1 Production: The Impact of Zero Setup Time
The ultimate ideal of manufacturing is "Lot 1 production." Efficiently manufacturing products with different specifications one by one. Producing at a cost comparable to mass-produced goods while meeting individual customer requirements.
Conventionally, this was difficult.The reason lies in "setup changes." Every time product specifications change, it is necessary to reprogram the robot, replace jigs, and change inspection standards. This setup change takes from several hours to several days.
Physical AI is beginning to realize "zero setup time."
The key is "automatic recognition" and "automatic adjustment." The AI recognizes the product with a camera."This is specification A."It makes this judgment. Then, it calls up the assembly procedure for specification A learned in advance. It automatically adjusts the robot's operating parameters.
If the next product is specification B,the AI recognizes "This is specification B"and switches to the procedure for specification B. Human intervention is unnecessary. Changeovers are completed in seconds.
As a practical example, there is a mid-sized machinery manufacturer in Germany (anonymous). This company manufactures industrial pumps. The specifications for the pumps differ for each customer. There are thousands of combinations of flow rates, pressures, materials, and configurations.
Conventionally, this company used batch production. They would make 10 units of the same specification at once, then switch to the next specification. It is efficient, but delivery times are long. Customers are kept waiting for several weeks.
In 2023, this company introduced collaborative robots from Universal Robots (described later) and an AI vision system. The robots recognize the parts and automatically select the assembly procedure.
As a result, lot-size-one production became possible.Once an order is received, manufacturing can begin immediately. Delivery times were shortened from several weeks to several days. Customer satisfaction improved significantly, andorders increased by 20%.
The impact of "zero changeover time" changes the business model of the manufacturing industry. The advantage of mass production is fading, and the value of customization is increasing. An era is arriving where optimal products can be efficiently provided to each individual customer.
Figure 2: Realization of zero changeover time
2-2. Learning-based collaborative robots from Fanuc (Japan) and Universal Robots (Europe)
Fanuc (Japan) and Universal Robots (Denmark) are leading the way as robot manufacturers realizing high-mix assembly lines.
Fanuc is a Japanese company with the world's top market share in industrial robots. It is known for its yellow robots. Fanuc's strengths are "precision" and "reliability." Its durability, which allows for continuous operation 24 hours a day, 365 days a year without failure, is highly regarded.
Since around 2020, Fanuc has accelerated the integration of AI technology. It has built an IoT platform called "FIELD system (Fanuc Intelligent Edge Link and Drive system)" to integrate data from the entire factory.
Fanuc's learning-based robots demonstrate their power in "bin picking." Bin picking is the task of picking up parts that have been randomly placed in a box. The position and orientation of the parts are also irregular.
With conventional robots, 3D models of parts had to be registered in advance and recognized through pattern matching. However, every time a new part arrived, a 3D model had to be created.
Fanuc's latest system uses AI for automatic learning. It learns the features just by being shown a few dozen parts. Creating 3D models is unnecessary. The time required to handle new parts has been shortened from several days to several hours.
On the other hand, Universal Robots (UR) is a manufacturer specializing in collaborative robots. "Collaborative robots" are robots that can work safely in the same space as humans. Conventional industrial robots needed to be enclosed by safety fences.
UR's robots have built-in force sensors. They stop immediately if they bump into a person or object. They are also lightweight and easy to install. The price (from 3 million yen), which makes them easy for small and medium-sized enterprises to introduce, is also attractive.
UR has built an ecosystem called "UR+." Third-party AI vision systems, grippers (holding devices), and software can be easily integrated. Customers can choose the optimal combination to suit their own applications.
For example, a small Japanese parts manufacturer combines a UR robot with Covariant's AI vision. The UR robot grabs the parts, and the Covariant AI performs quality inspections. This flexibility enables high-mix support.
Table 3: Fanuc vs. Universal Robots Comparison
2-3. Case studies of realizing safe collaboration with workers
An important value of physical AI lies in "safe collaboration with humans."It is not about full automation, but about humans and robots working together.
Let's look at the case of a Japanese automotive parts manufacturer (anonymous). This company introduced collaborative robots to its engine parts assembly line.
Assembly work is complex. More than 20 parts are assembled in a precise order. Some tasks require skill. Some tasks are simple but physically demanding.
This company clarified the division of roles. The robots are responsible for "transporting heavy parts" and "repetitive screw tightening." Humans are responsible for "precise positioning" and "final inspection."
What is important is that the robot "understands" human movementsIt uses AI vision to recognize the worker's position and movements. If a worker approaches, the robot slows down. If a worker reaches out, the robot waits.
This "tacit coordination" increases efficiency. Workers treat the robot not as a "tool" but as a "colleague." If they signal with their hand to "wait a moment," the robot waits. If they point and say "this one next," the robot brings the corresponding part.
The introduction effect was remarkable. Productivity increased by 25%. At the same time, the physical burden on workers was reduced. The number of workers complaining of lower back pain or stiff shoulders was halved.
The evaluation from workers is also high. One worker says, "I was anxious at first, but now I would be in trouble without it." Another worker is happy, saying, "My body feels easier, and I can concentrate on work that requires thinking."
There are three keys to safe collaboration.
First, "predictable movement."If a robot moves suddenly, humans are surprised. The robot signals what it will do next using LEDs or sound.
Second, "flexible stopping."If a human says "stop" (or signals with their hand), it stops immediately. There is no need to press an emergency stop button.
Third, "learning relationships."The robot learns the habits of this worker. "This person is right-handed, so let's place parts on the right side." "This person prefers a fast pace, so let's move a little faster." It adapts to individual workers.
3. Analysis of domestic and international introduction cases
3-1. Covariant (USA): Automation of logistics warehouse picking
Regarding Covariant, which was introduced in Episode 6, let's look at a specific introduction case here.
A major US daily necessities e-commerce site (anonymous) introduced Covariant's system in 2022. This company handles tens of thousands of products, from toothbrushes to detergents.
The challenge was the diversity of products.Size, shape, weight, and material are all different. There are soft sponges and hard plastic containers. There are products in transparent bags and products in black boxes.
Human workers identify these through experience and grab them appropriately. However, the labor shortage was serious. During busy seasons, people could not be gathered even by raising the hourly wage.
Covariant's system was first introduced as a pilot. One robot was installed at one picking station. For three months, AI was trained while collecting data.
The initial success rate was about 85%. Out of 100 picks, 15 failed. However, through Covariant's fleet learning, the AI evolved daily. After three months, the success rate reached 95%.
The company, having confirmed the success, decided on a horizontal rollout. They added 10 more robots. And one year later, they added another 50.
Currently, in this company's warehouse, robots are responsible for 60% of all picking tasks. The remaining 40% is handled by humans for products that robots are not good at (extremely large, fragile, etc.).
The effect is clear. Processing capacity improved by 40%. The labor shortage was also resolved. Even during busy seasons, additional recruitment is no longer necessary. The ROI is expected to be recovered in 2.5 years.
What is important is "phased introduction" and "continuous learning."Do not introduce it fully all at once; verify on a small scale, learn, and expand. This approach was the key to success.
3-2. Dexterity (USA): General-purpose support for high-mix products
As a case study of Dexterity, let's look at the collaboration with the US logistics giant FedEx.
FedEx's distribution centers handle a huge variety of packages. From envelopes containing documents to large boxes of home appliances. The shapes and weights are completely different.
Conventionally, sorting work was done by humans. They take packages from a conveyor belt, check the destination, and sort them into the corresponding delivery truck. It is simple, but physically demanding. Especially during busy seasons (such as the Christmas season), workers suffering from lower back pain appeared one after another.
In 2023, FedEx introduced Dexterity's RaaS system. Since it is RaaS (Robotics as a Service), initial investment is unnecessary. You can use the entire robot system for a monthly fee.
Dexterity's DexR system integrates multiple robot arms and AI vision. When a package is placed on a conveyor belt, the AI identifies it instantly. It judges size, weight, and destination, all in a few seconds.
The robot arm grabs the package with appropriate force. Light envelopes are handled gently, and heavy boxes are handled powerfully. Then, it puts them into the container of the corresponding delivery truck.
The processing speed is about 1.5 times that of humans. Furthermore, it can operate 24 hours a day. Securing personnel for night shifts was difficult, but with robots, that is not a problem.
Following the success at the pilot site, FedEx decided to deploy it to major distribution centers across the United States. Hundreds of DexR systems are scheduled to be introduced.
Dexterity's strength is "end-to-end support." It does not just pick; it consistently performs inspection, packing, and labeling. This one-stop nature is highly valued by customers.
3-3. ROI analysis: Quantitative evaluation of introduction costs and effects
Based on the cases so far, let's analyze the ROI quantitatively.
The introduction cost of a typical picking robot system is as follows.
Hardware (robot arm, camera, sensor, etc.): approx. 3 million yen/unit. Software license (Covariant, etc.): initial cost 2 million yen + 200,000 yen/month. Installation/integration cost: approx. 2 million yen. Total initial investment: approx. 7 million yen/unit.
Annual operating costs: Maintenance 500,000 yen + Electricity 200,000 yen + Software 2.4 million yen = approximately 3.1 million yen.
Let's look at the benefits side. Annual labor cost for a picking worker (USA): 2,500 yen/hour × 8 hours × 250 days = approximately 5 million yen. Assume one robot does the work of 1.5 workers.
Annual benefit: 5 million yen × 1.5 workers = 7.5 million yen (labor cost reduction). Annual cost: 3.1 million yen. Net: 4.4 million yen/year. The initial investment of 7 million yen can be recovered in approximately 1.6 years.
However, this calculation only considers 'labor cost reduction.' In reality, there are other benefits: increased sales due to higher processing speeds, fewer complaints due to improved quality, and higher employee satisfaction due to better working environments. Including these, the ROI becomes even better.
Conversely, if the same system is introduced in Japan, where labor costs are about half (1,200 yen/hour), the ROI recovery period will be about 3 years. Even so, it is well within 5 years, making it sufficiently profitable.
Figure 3: ROI Calculation Model (10 units introduced)
Chart 4: ROI Calculation Example
Summary
In Part 8, we looked at specific use cases in assembly, picking, and processing. Handling of irregular objects has been realized through the fusion of vision AI and tactile sensors. Foxconn's electronics assembly, food cutting operations, and apparel sorting were impossible just a few years ago.
High-mix assembly lines have been realized with zero setup time. Learning robots from Fanuc and Universal Robots are enabling lot-size-one production. Safe collaboration with humans has also been demonstrated on the factory floor.
Logistics warehouse cases from Covariant and Dexterity show that ROI can be recovered in 2-3 years. Phased introduction and continuous learning are the keys to success.
Next time, in Part 9, we will look at use cases in inspection, maintenance, and logistics. How are AI-based visual inspection, predictive maintenance, and autonomous mobile robots changing the workplace? We will deliver further examples and insights.
Next Issue Preview
Part 9 (Use Cases 2) 'Inspection, Maintenance, and Logistics—Precision Beyond Human Vision and AI Predictive Maintenance'
Physical AI is powerful not only in assembly but also in inspection, maintenance, and logistics. AI-based visual inspection achieves precision that exceeds the human eye. It does not miss even a 0.1mm scratch. Predictive maintenance predicts equipment failures in advance, minimizing downtime. Autonomous mobile robots move freely within the factory and operate 24 hours a day. In Part 9, we will clarify how these technologies are implemented, what the ROI is, and how Japanese companies should utilize them.
References and Sources
1. Foxconn Technology Group (2024) 'AI Robotics in Electronics Assembly', Foxconn Industrial Internet
2. Fanuc Corporation (2024) 'FIELD system and AI-Powered Bin Picking', https://www.fanuc.co.jp/
3. Universal Robots (2024) 'Collaborative Robots and UR+ Ecosystem', https://www.universal-robots.com/
4. Covariant Inc. (2024) 'E-commerce Warehouse Automation Case Study', Covariant Customer Success
5. Dexterity Inc. (2024) 'FedEx Partnership: Logistics Automation at Scale', Dexterity Press Release
6. IEEE Robotics & Automation Magazine (2024) 'Vision-Force Fusion in Robotic Manipulation', IEEE RAM
7. Journal of Manufacturing Systems (2024) 'Lot-Size-One Production with AI-Enabled Robots', Elsevier
8. International Journal of Production Research (2024) 'ROI Analysis of Physical AI in Food Processing', Taylor & Francis
9. Robotics and Computer-Integrated Manufacturing (2024) 'Safe Human-Robot Collaboration: Best Practices', RCIM Journal
10. McKinsey & Company (2024) 'The Economics of Physical AI: Cost-Benefit Analysis Framework', McKinsey Manufacturing
Keywords
#IrregularObjectHandling, #VisionAI, #TactileSensor, #Foxconn, #HighMixLowVolumeProduction, #SetupChange, #LotSizeOneProduction, #Fanuc, #Universal_Robots, #CollaborativeRobot, #Covariant, #Dexterity, #LogisticsAutomation, #Picking, #ROIAnalysis, #PhysicalAI, #AI, #IoT,
About the Author
Tatsuhiko Hatakeyama CEO, DCTA Inc. / Manufacturing and Environmental Consultant
After working in plastic development and factory design, operation, and management at a major chemical company (Mitsubishi Chemical), he is now a consultant based in Shonan and Yokohama, working on environmental protection and manufacturing improvement activities. With extensive experience in recycling, CO2 reduction, and PFAS countermeasures, he supports manufacturing companies aiming for decarbonization and carbon neutrality in building smart factories using IoT, AI, and digital twins. His activities extend beyond Japan to Southeast Asia, the Middle East, India, Europe, and the United States, developing international initiatives to build a sustainable future together. As a NewsPicks Expert, he received the 'EXPERT AWARD 2024 Best Flash Opinion Category'. On note, he writes serials on diverse themes such as environmental regulations, PFAS, the Cyber Resilience Act, and time-machine management, disseminating practical information for manufacturing practitioners.
note: https://note.com/iceman_23
Website: https://www.dctainc.com/
Contact: t_hatakeyama@dctainc.com