For all the talk around humanoids and robots entering warehouses, factories and homes, one of the more immediate problems in physical AI lies behind the machines. There is not enough diverse real-world data to train them.
Instawork expects to facilitate about 20 million hours of robotics-related work across its ecosystem in 2026, according to Vaibhav Pandey, Country Head and Senior Director, Product, Instawork. Yet Pandey says that would amount to less than 0.04% of the data the company estimates will eventually be required.
The scale of that gap is also beginning to influence the mandate of Instawork’s India operation. The company entered India nine years ago as an engineering and product design talent hub. It has since expanded its India workforce to about 6,000, with the country accounting for roughly half of its global workforce, he said. The India entity is now also involved in work around physical AI, including the collection of real-world data and training workers for robot-related operations.
“We started as a talent hub for engineering product design, like a typical GCC. And then, over the last one year, we’ve started launching our operations in the market. So we are no longer a GCC; we are operating as a unit that networks and operates in India,” Pandey says.
The shift, as described by the company, points to a GCC mandate that is moving beyond a conventional engineering and talent function towards work connected to an emerging business capability.
The data has to come from the physical world
Instawork’s focus is on physical work that involves hand movements, manipulation and other forms of activity that generate economic value. The objective, adds Pandey, is not to collect data simply for the sake of building a larger dataset, but to capture work that robots could eventually be expected to perform.
“The intention is not to collect data for the sake of it, but to collect natural, valuable work that happens in the real world,” he says.
The distinction is important to the company’s approach to physical AI. The Instawork team explains that conventional AI models can learn considerably from text and existing images, whereas physical systems need information about how objects and environments behave during interaction.
There is, for instance, no straightforward way to describe through text what it feels like to hold and manipulate an object. The company therefore sees video and sensor-based data from physical activity as necessary to help models learn aspects of physical interaction.
That also limits the usefulness of simply recording people performing routine work. The data needs to capture the movements and physical manipulation relevant to the task a robot is expected to learn.
The problem is diversity, not only volume
The discussion around robotics data often begins with the question of how much is required. The Instawork team argues that the more difficult issue is the range of activities represented in that data.
Robots have long been deployed for specialised industrial tasks. The difference with physical AI, as described in the interaction, is the attempt to develop foundation models capable of performing a wider range of activities rather than training a system for one narrowly defined workflow.
“The gap is in diversity,” the team points out. “On top of scale, we need diversity in how many different kinds of actions and jobs and tasks you can cover. Because that will make the model truly intelligent so that they can transfer their learning from one use case to the other.”
Pandey adds that the company does not expect the data-collection requirement to reach a point where all possible information has simply been gathered. As robots take on different tasks, he said, the data requirement will also change.
“It will never end where you will say, ‘Okay, now I have collected all the data possible to empower my foundation for all robots,’” he pointed out.
He compares the potential evolution with the development of large language models, where increasingly specialised applications created continuing requirements for training and fine-tuning. In physical AI, a humanoid capable of one set of tasks could eventually be expected to perform substantially different activities, creating another requirement for data.
A workforce around the robots
The company’s account also places workers inside, rather than outside, this transition.
Pandey reveals about 20,000 workers have been certified and given formal training related to robot operations and deployment. He described these activities as robot-rangling services, where workers can work alongside robots, help deploy them and assist with their maintenance.
He says the wider data-collection activity involves thousands of workers and businesses, with workers taking up projects according to their availability.
Pandey also frames the impact of robotics on physical labour in terms of changing work rather than simply eliminating it. He points to previous technology shifts, including tractors, industrialisation and the IT revolution, as examples of technologies that altered how work was performed.
“What technology basically does is it changes how you work, changes how you live,” he says.
Whether the new roles around robots develop into a substantial employment category is not established by the current trajectory. For now, the company’s description suggests that the training and deployment of physical AI systems themselves require human workers.
From talent centre to market-facing operation
The change in mandate is also visible in how Pandey described the India entity.
He says Instawork began in India with five or six people, mostly engineers, before expanding into product, design, HR and other functions. The India operation is now a wholly owned subsidiary and works with businesses and workers in India. Its relationships also include robotics companies based in India and the US.
That is a different operating responsibility from simply providing engineering talent to a global headquarters. In the company’s description, the India operation is now participating in activities that require access to local workers and businesses as well as technology and data capabilities.
Physical AI itself remains at an early stage. Pandey says humanoids are already operating and moving around physical environments, but are not yet capable of highly specialised activities involving specific equipment. He describes the timeline for wider deployment in areas such as warehousing, supply chains and manufacturing as difficult to predict.
The uncertainty extends to the data requirement. The 20 million hours cited for 2026 may represent a substantial operational effort, but Pandey’s own estimate places it at less than 0.04% of what will eventually be needed.
For Instawork’s India operation, that leaves the immediate task less about declaring the data problem solved than about building the capacity to keep addressing it. And in doing so, the company is describing an India mandate that has moved from engineering and product development towards an operating role around a technology that is still being trained to understand the physical world.