.png)
Have you ever been to Home Depot?
Why go to Home Depot instead of a retailer like Costco that sells a bit of everything?
- You go to Costco when you want groceries, household supplies, and other products in one trip.
- But if you’re building a deck over the weekend, you’re more likely to go to Home Depot. The same is true if you’re a carpenter or plumber buying supplies for a job.
So, why hasn’t a generalist retailer like Costco completely replaced a specialized retailer like Home Depot?
Because Home Depot offers something Costco doesn’t. Home Depot offers specialized parts for home improvement, construction, and electrical projects that wouldn’t make sense for Costco to stock. It has employees with trade experience who can help with projects, a pro desk for repeat professional buyers, and infrastructure for handling long lumber and tubing that Costco doesn’t support. It serves a different set of customer needs than Costco. Generalizability isn’t the only dimension that matters.
Today, Home Depot has a $300B market cap. Costco is a $400B company. Home Depot has not been eaten alive by Costco. In fact, a Wall Street Journal analysis identified Home Depot as the highest-returning U.S. stock over the 45 years following its IPO.
A common misconception in physical AI is that once a generalist policy can execute the motions required for any task, robotics is “solved.” The assumption is that the companies building those policies will then capture the entire robotics market. This is similar to assuming that a retailer selling everything is inherently better than one serving a particular domain. But as we’ve discussed, these retailers provide different value to different customers.
In the physical world, executing a motion is not sufficient to deliver return on investment (ROI). Generalizability is only one factor, and it can come at the expense of others, such as throughput. Just as Home Depot can coexist with Costco, specialized robotics companies can serve requirements that generalist systems aren’t built around.
What industrial robotics buyers need
Robotics buyers consider several factors when evaluating a system:
- ROI
- Throughput
- Quality and precision (i.e., in the food industry, placement quality and spillage)
- Reliability
- Yield savings
- Food safety (specific to food)
- Worker safety
- Ease of use for non-technical users. If a robot requires a dedicated operator, that affects ROI.
Our food manufacturing customers tell us that a worker typically performs one to three tasks throughout the day, often just one. They don’t need a system that can assemble food, fold laundry, and pick up toys if that flexibility comes at the expense of throughput, reliability, and yield savings.
We at Chef focused on food manufacturing rather than restaurants because production volumes are much higher and jobs are more specialized. There are dedicated teams that handle preparation, assembly, and sanitation. An assembly worker spends the day doing one job: assembling food. Even within assembly, however, variability remains significant. Manufacturers need domain generalizability. For Chef, that means building robots that can handle different ingredients, portion sizes, trays, compartments, and conveyors.
Generalizability matters, but it differs between generalizing within a domain and generalizing across unrelated tasks. A food assembly robot needs to adapt to the variability of food production. It doesn’t need to fold laundry. The same is true in automotive factories, where a worker may handle wire harnesses all day, or in other manufacturing facilities, where a worker may spend an entire shift kitting parts. Home Depot follows the same principle: it offers a broad range of products within a particular domain.
Why throughput matters
Throughput is one of the most important factors in industrial robotics. Our customers’ production lines often run at 30 to 45 trays per minute. That means a tray moves through the line approximately every 1.3 to 2 seconds.
Our customers deploy robots to replace manual work and deliver ROI. To provide that return, the robot needs to outperform a person. In other words, customers need superhuman performance. This is difficult to achieve with techniques commonly used for generalist models, such as imitation learning from human demonstrations. Behavior cloning is simply too slow for our application and most industrial applications.
We’ve learned that food manufacturers are extremely sensitive to ROI. They may have 50 potential projects, each with an expected return, but will pursue only one, two, or three. They prioritize projects with a clear business case and a meaningful impact on metrics such as throughput and revenue. Manufacturers don’t choose a system based on whether it uses imitation learning, foundation models, world models, world action models, or hybrid AI. They evaluate it by the throughput, reliability, quality, and yield it delivers.
Different requirements for robots at home
A general-purpose robot can make sense in the home for several reasons:
- High task diversity: A household robot may need to fold laundry, prepare food, pick up toys, and perform many other tasks. Task generalization matters because the work changes throughout the day.
- More flexibility in throughput: If folding laundry takes 30 minutes instead of 20, that may be acceptable.
- More tolerance for imperfect results in some tasks: An imperfectly folded shirt 60-70% of the time may be acceptable, but an incorrectly assembled meal on a production line is not.
Generalist models make sense in high-diversity, high-mix, low-volume applications like the home. However, in industrial settings, where robot performance and hardware costs already make strong ROI possible, generalizability is only one factor. It cannot deliver ROI on its own.
Consider another analogy. A fish is better suited to swimming than a monkey, and a monkey is better suited to climbing trees than a fish. A human can swim, climb trees, make bagels, and cut lumber. In our experience, food manufacturers want the equivalent of fish and monkeys: robots that perform a specific task exceptionally well and deliver strong ROI. Homeowners may want the equivalent of humans: robots that can perform a wide range of tasks.
An industry-by-industry approach to physical AI
We believe that while robots will eventually enter the home, today, however, much of the work in America’s factories, warehouses, and other industrial settings is still done by people. We expect robots to automate these industries before they automate the home.
For industrial applications, we believe an industry-by-industry approach is necessary. Each industry needs a solution that performs exceptionally well and generalizes within that domain. In construction, that might mean an excavator with dust-resistant hardware and policies designed for construction tasks. In surgery, it means surgical-grade hardware and policies capable of extremely fine, dexterous manipulation. In automotive manufacturing, it might mean a robot that can handle a wide variety of wire harnesses.
Our technical approach is to use whatever maximizes throughput, yield, and uptime, including foundation models. Our advantage comes from the domain we focus on, the robots we deploy in production, and the performance feedback we use to improve them.
Want to learn how Chef robots can support your production needs? Contact our team to discuss your throughput, safety, and ROI requirements.
.avif)
