Author: Sumir Meghani, Instawork Robotics Labs (IRL); Compiler: Peggy, BlockBeats

While most people are still discussing "whether robots will replace human jobs," this article believes that not only will humans not be replaced, but they are becoming an indispensable key infrastructure in the "physical AI system."

The core bottleneck of the current industry does not lie in algorithms or hardware, but in "data and implementation capabilities." Robots need to be trained by observing skilled humans operating in real environments, but high-quality, diverse physical world data is extremely scarce, resulting in the so-called "100,000-year data gap." This has also brought to the fore a long-neglected capability—skilled, dispatchable, and verifiable human labor.

Under this framework, the role of human beings has been re-disassembled: they are not only the "data source" for training machines, providing standardized and labelable operation processes; they are also the "field nodes" that support the operation of the system, responsible for maintenance, repair and remote control; and finally enter a "human-machine collaboration market" connected by a platform, becoming a necessary condition for the large-scale implementation of robots.

In fact, technological change will not eliminate labor, but reconstruct the division of labor. From ATMs to the Internet, every technological leap is accompanied by anxiety about employment, but what is often changed is not "whether there is a job", but "how work is redefined." In this round of technology cycles represented by humanoid robots, the same path is repeating itself: tasks are dismantled, capabilities are standardized, positions are reorganized, and new careers are created.

The real opportunity lies not in "replacing people," but in who can build that bridge and transform human capabilities into scalable data, operation and maintenance systems, and collaboration networks.

The following is the original text:

A year ago, I asked a question that was perhaps somewhat unusual for the labor market: What will happen to the “Pros” on our platform when the robots arrive?

Our vision is to create economic opportunities for Pros and partners around the world. Today, more than 10 million Pros depend on us for their livelihoods, and many of them are already thinking the same question. We have a deep responsibility for this and must give answers.

At the same time, we also observed an unexpected phenomenon: some robotics companies have begun to appear on our application platform to cooperate with our Pros. They need people with expertise in robot training tasks, but also in the diverse business scenarios where robots will be deployed in the future. And what they rely on is the labor system we have been building.

At that moment, it suddenly became clear: Instawork could provide human labor for the “physical AI economy.”

"One Hundred Thousand Year Problem"

Ken Goldberg summarized this problem as the "100,000-year data gap": on the one hand, there are massive amounts of data used to train language models; on the other hand, there are extremely limited and highly specialized data used to train robots to complete precise operations in the physical world.

Note: Ken Goldberg is a very influential scholar in the field of robotics and artificial intelligence, as well as an artist and interdisciplinary researcher

It's this gap that means that despite the billions of dollars pouring into robotics companies, we still haven't seen humanoid robots cleaning rooms in hotels or unloading goods in warehouses... at least not yet.

Our estimate is that the industry as a whole will collect approximately 100,000 hours of training data in 2024; by 2025, this number will grow to 1 million hours; and by 2026, it is expected to reach 20 million hours. That’s exponential growth, but even that only accomplishes 0.04% of bridging the gap.

More and more companies are joining the race to build humanoid or general-purpose robots: Basic Model Labs is developing vision-language-action (VLA) models, hardware companies are building physical machines, and players in the middle are also emerging. Capital investment has reached tens of billions of dollars. All these participants face the same bottleneck: data.

But the key is that we have actually seen this scene before.

When automated teller machines (ATMs) emerged, almost everyone predicted that bank tellers would disappear. But the result was just the opposite—the number of tellers increased. ATMs have reduced branch costs, allowing banks to open more branches; and the role of tellers has shifted from counting money to maintaining customer relationships.

This pattern recurs with every major technological change: the Industrial Revolution, electrification, the Internet. New technologies do not eliminate jobs; they reshape them and create new opportunities.

A new wave is coming, but this time, it looks more like us: with arms, legs, and eyes.

A three-act drama of physical AI

Act 1: Training the robot

Over the past year, I’ve reached out and asked for advice from some of the best in robotics learning around the world—from researchers and lab leaders to entrepreneurs building dexterous robotic hands and even complete humanoid robots. I was impressed by their generosity in sharing their time and insights. Frankly speaking, we were not originally part of this industry; but the more I listened, the clearer I saw the space that Instawork could enter.

One idea that comes up again and again is that robots learn by watching skilled humans perform delicate physical tasks in real environments. This means everything from chopping vegetables to standard knife skills, to navigating a crowded warehouse, to making hotel beds to brand standards. The problem is that collecting this kind of data at high quality is extremely difficult—you can’t just stick a camera on someone and start recording. The data must cover a diverse range of environments, tasks and hand movements; more importantly, the people performing these tasks must be truly expert. Otherwise, a robot trained with "bad knife skills" will only learn "bad knife skills" (which is not a good thing for anyone).

This is essentially a workforce operations problem: how to recruit skilled workers, train them, ensure the quality of output, and manage a distributed workforce network in different geographies and scenarios - these are exactly what we have been doing. We have more than 10 million Pros with proven skills, covering hundreds of task types; we have established in-depth relationships with partners and are able to enter real business scenarios; and we have data on who can maintain stable attendance and continue to complete work with high quality. This combination cannot be copied from scratch by any data collection company. In fact, many laboratories have already approached us on their own initiative, and now we are cooperating with most of the leading teams in the field.

Act 2: The Rise of the Robot "Trainer"

One thing is often overlooked: robots also need people.

An executive at a leading robotics company told me that they had a critical component that needed to be replaced every 4–6 months—not often enough to have a dedicated technician, but enough that downtime would be a significant cost. With the popularity of autonomous driving, delivery robots and various types of automation deployment, more and more companies are facing similar problems: expansion requires on-site support, but it is not economically feasible to have dedicated personnel in each market.

We have launched pilot projects with a number of robotics companies, covering services such as battery replacement, parts replacement and robot maintenance. At the same time, we have established a robot certification system for hourly employees—which can be said to be the first attempt in the industry. In the first few weeks alone, more than twenty thousand Pros have been certified.

On the data collection side, certified Pros learn how to operate wearable cameras, capture high-quality video, and annotate sensor data—when the robotics lab needs to record hours of bed-making processes in real hotel suites, they get professionals, not newcomers who are "learning by doing." On the technical support side, Certified Pros master hardware diagnostics, safety regulations, and maintenance procedures for specific robotic systems.

Imagine a scenario where a logistics company deploys a fleet of automated robots in more than a dozen warehouses. At two in the morning, a robot in a Memphis warehouse has a navigation error, or a piece of equipment in Phoenix needs a sensor module replaced. Instead of waiting days for a factory technician to fly to the site, a certified Instawork Pros can arrive and fix the problem within hours. At the same time, we are also developing remote control training based on VR to support the laboratory to break through the limitations of simple on-site recording after the scale of data collection is expanded.

If billions of AI devices will be deployed over the next decade, the opportunity lies not just in maintaining them, but in creating entirely new job categories: robotics technicians, fleet operators, remote control specialists, and even a new job title we have yet to name.

Act 3: The market for human-machine collaboration

Last year, I had lunch with the CEO of one of the world’s largest hotel groups. They are thinking hard about how to improve the consistency of housekeeping through automation. A large number of robotics companies want to deploy products in their hotels, but it is difficult for them to judge what is just a "demonstration effect" and what is a real "operational result." And we are very familiar with these scenarios, processes and pain points - because we have already provided services in these places.

We are building a "robot service marketplace" - connecting robotics companies with enterprises ready to deploy automation. We already serve both supply and demand sides, which means that we are not just "matching up", but can actually promote implementation.

The future is not "robots replacing humans", but "robots and humans collaborating". That’s exactly what Instawork Robotics Lab is all about: three capabilities, one platform – training robots, enabling them to operate in the real world, and connecting them to the business scenarios that really require them.

Bridge

In every major technological change, the question is never whether new jobs will be created—the answer is always yes. The real question is: who will build the bridge between the present and the future.

We believe skilled humans are needed at every stage of the process—from training the first generation of robots, to deploying large-scale systems, to designing the human-robot collaboration processes of the future. We hope that Pros on the platform will be present throughout the entire process.

In the "Physics AI Revolution", Instawork hopes to be the bridge: accumulating deep experience in the most influential industries; already providing training data for robotics laboratories; already cultivating certified talents for data collection and field operations; and also building a market that connects robots with enterprise needs.

We are looking forward to the next phase.