The next generation of robots is running into a basic shortage: there is no web-scale archive of people folding laundry, packing boxes or manipulating tools that machines can download and study.
That gap is pushing physical-AI developers to build their own data supply chains. TU News, summarizing Financial Times reporting, describes a mix of purpose-built robot gyms, paid task-recording programs and data facilities in lower-cost labor markets. The goal is to collect enough examples of physical work for models to learn how perception and movement fit together.
Why web data is not enough
A language model can train on text and images gathered from the internet. A useful robotics system needs additional information about depth, movement and the amount of force required to complete a task. Video helps, but developers also need demonstrations connected to hardware and sensor data.
Companies including Generalist, Physical Intelligence, Neura and Skild are using outside providers or building robot gyms. In those facilities, workers control robotic arms through tasks that machines can later attempt to reproduce. Other companies are sending cameras and recording equipment into factories, offices, homes and vacation rentals to capture first-person footage.
Simulation and “world models” can supplement that material, but robotics developers are still spending heavily on physical data collection. PitchBook data cited by the Financial Times put year-to-date fundraising for robotics and physical-AI companies at nearly $48 billion.
Turning people’s tasks into training sets
Figure, a humanoid robotics company, outlined a program that pays people to record themselves doing tasks at home and work. The FT reported that it had about 44,000 active participants earning an average of roughly 94 cents per task. The company also planned to spend more than $1 billion over 12 months on services and computing power for training-data generation.
Physical Intelligence is testing a different part of the problem: whether models can adapt to unfamiliar environments. The startup has run robots in several San Francisco rental homes rather than keeping every test inside a fixed lab.
The data industry is following the spending. Scale AI has opened a robotics data facility in Mexico City, while Xdof operates facilities in Shenzhen and Jakarta. Xdof CEO Philipp Wu said the labor cost and volume of data required made equivalent work in the United States uneconomical.
Those economics create a labor issue alongside the technical one. The FT found factories in Gurugram, India, where workers were asked to wear recording equipment, sometimes without extra pay. The resulting footage can become a commercial training asset even though the workers may be doing their normal jobs.
More data is still a hypothesis
Robotics companies are betting that a version of the scaling pattern seen in language models will apply to physical AI. Dyna reported stronger model performance when it used as much as 1 million hours of first-person video.
That does not settle the question. UC Berkeley professor and Ambi Robotics co-founder Ken Goldberg warned that generating large quantities of data is no guarantee of useful results. Before ordering millions of repetitions, developers still have to determine which tasks, environments and demonstrations actually improve a model.
For now, the robot gym is both training ground and experiment. Companies know they need more real-world experience than the internet can provide; they are still learning what kind of experience is worth collecting.
Comments