Private Data Will Shape the Future of AI

When people think about warehouses, they imagine big machines, conveyor belts, and thousands of boxes moving smoothly every day. When people think about robotics, they imagine smart robots learning how to move like humans, folding clothes, assembling electronics, or packing shipments.

But most people do not think about what happens in between. They do not think about the data. And they definitely do not think about how messy, invisible, and broken that data can be.

I know because I lived it.

This year, I decided to work warehouse jobs on purpose. I already had a background in IT and AI, but I wanted to understand physical labor. I took night shifts at meal kit fulfillment centers. I operated pallet jacks. I loaded trucks. I packed boxes until my hands hurt. I did inventory audits. I wrapped pallets. I experienced the full chaos that happens when human speed, mistakes, and tiredness meet tight deadlines.

And something kept bothering me. How is it that in places with hundreds of cameras recording everything, mistakes still happened every day? Why was nobody using all that video to help? Why were we still working in the dark?

At first, I thought maybe managers just were not paying attention. But then I started asking questions. I learned that most warehouses do record video footage, but it is mainly used after something goes wrong, like checking tapes after a missing item, a crash, or a theft. The video is not labeled, not organized, and definitely not used to help workers in real time.

I also remembered something from when I first signed up to work. I had signed a photo release agreement that allowed the company to use pictures of me for marketing or training. But I never signed anything clear about how my video footage would be used. I thought about exercising my data rights, asking for copies of all videos showing me, but realized it would be very difficult. Not only because of cost, but because the footage would also show other workers, which brings up privacy and legal risks. And in states where there are no strong privacy laws yet for employees, there is almost no clear path for workers to access their own data.

One day, I had a bad idea. What if I filmed everything myself? I thought about buying smart glasses with a camera to capture stacking errors, messy loading zones, missing labels, and broken pallets to train my own model. But I quickly realized that would be unethical. I would be invading my coworkers’ privacy, even if my intentions were good. I would be contributing to the same problem I wanted to fix.

That was the moment I understood something important. It is not enough to have data. If we cannot label it, protect it, or use it responsibly, it is almost worthless.

As I kept working, I also noticed something else. Companies are willing to use data, but only when it benefits them directly.

When I rode with a delivery driver one Friday morning, I noticed a small camera retrofitted inside the truck. It was not part of the original design. The camera would beep or alert the driver if he went over the speed limit, used his phone or scanner while driving, or even yawned too hard. If the driver yawned too much, the system would recommend pulling over to rest. I realized companies are willing to invest in data systems when it helps them reduce insurance costs, prevent lawsuits, or meet safety regulations.

In the warehouse, I saw something similar. New hires were given Verve exosuits to help with lifting posture. The suits tracked motion and collected data on how well workers lifted heavy boxes during their first few weeks. Again, the goal was clear: reduce injuries, cut costs, and manage risk.

But when it came to quality control, inventory accuracy, or worker support, the same level of data investment was missing. Companies focused more on protecting themselves than on improving the daily work experience.

At the same time, I noticed another important thing.

Some workers were worried about what technology might mean for their jobs. One co-worker joked, "Thank God the exosuit can't lift for us yet." Another said, "It will take our job," when I pointed out how easy it would be to install a system that checks pallet quality before loading shipments.

These conversations showed me that warehouse workers are not blind to change. They know when technology is designed to support them, and they know when it is moving toward replacing them. Augmentation and automation are not the same thing, and workers can feel the difference even when no one says it out loud.

Robots do not learn in a vacuum. They need data, lots of it, to learn how to move, how to handle mistakes, and how to adapt.

Today’s robotic foundation models, like the ones being developed at places like DeepMind, Stanford, and now startups like Physical Intelligence, depend on large-scale demonstrations of real tasks. Folding laundry. Packing electronics. Assembling parts.

But here is the hidden truth. Most real-world warehouse data is too messy, too private, or too disconnected from the real human experience to be useful.

Without clean, ethical, labeled data from messy, fast-paced places like warehouses, robots will continue to struggle with general-purpose physical skills. The data they are trained on will be too simple, too idealized, or too unrealistic compared to the real world.

Here is what I believe we need to move forward.

We need a warehouse-focused data labeling platform. Something similar to what Scale AI built for autonomous vehicles, but designed specifically for warehouses. It would help companies label warehouse video data by task, such as stacking, wrapping, picking, lifting, and driving, with human oversight and machine support.

We need built-in privacy protections. Faces, tattoos, and identifying features should be automatically blurred by default. Workers should be notified about how video is used, and there should be clear ways to request deletion or correction if needed.

We need a new benchmark for physical warehouse tasks. A public or semi-public dataset, like an ImageNet or COCO but for warehouse operations, that captures real-world challenges: ripped boxes, bad lighting, blocked aisles, human improvisations. Something messy and real, not cleaned up for the cameras.

We need worker-centered ethics from day one. The goal should not just be to monitor people. It should be to build systems that assist them, prevent injuries, reduce errors, and make warehouse jobs safer and less stressful.

I did not learn any of this from reading papers or taking online courses. I learned it by clocking in at 6 PM, riding a pallet jack through freezing warehouses, lifting 80-pound boxes, and watching the small human moments that never make it into datasets.

There is a huge opportunity to help warehouses, workers, and robots get better together. But it starts with seeing the gaps, not just in technology, but in how we think about data, work, and dignity.

I am still early in my journey. I am still building FloorSight AI, trying to make small improvements. But I hope more people, from founders to researchers to operators, start asking these questions too.

Because the future of robotics does not just depend on better models. It depends on seeing the real world clearly, mess and all.