Robotics companies ramp up data collection as training bottleneck slows physical AI

Traders Union
All News
Financial News
Robotics ramp up data for AI

Highlights

Physical AI companies have raised nearly $48bn year to date, with a large share allocated to building bespoke robotics training datasets.
Mecka secured $60mn in funding from Nvidia and Sequoia, while Figure plans to spend over $1bn in 12 months on training data services and computing power.
Scale and Xdof are expanding robotics data collection to lower-cost markets like Mexico and Asia, where operator pay is about half the U.S. rate, to address economic pressures.

Costly push to build robotics training data

As reported by Financial Times, robotics labs are spending billions of dollars on equipment, computing power and data-gathering networks because physical AI models cannot rely on the vast web-based datasets that helped train large language models. Companies including Generalist, Physical Intelligence, Neura and Skild are using external suppliers and internal robot gyms, where workers operate bimanual robotic arms to perform tasks that machines later learn to emulate.

Some groups are also deploying cameras and recording kits to factories, homes, offices and holiday rentals to capture first-person, or egocentric, footage of people carrying out everyday work. Joe Fox Jr, director of robotics operations at Scale AI, says there is no equivalent of the internet for robotics developers to download a large training corpus from, turning access to high-quality data into a competitive race.

Investor appetite remains strong. Physical AI companies have raised nearly $48bn year to date, according to PitchBook data, and a significant share of that funding is going toward data generation. Mecka this week says it raised $60mn from investors including Nvidia and Sequoia, while Figure has about 44,000 active users recording tasks at home and at work and plans to spend more than $1bn over 12 months on services and computing power tied to training data.

Global expansion raises cost and labour questions

Data providers that previously served OpenAI, Anthropic and other AI labs are now opening robot gyms and data factories in the U.S., Mexico and parts of Asia, including China and Indonesia. Robotics models need a broader mix of video, depth perception and force-related data than text systems such as ChatGPT, Gemini or Claude, while developers are also using simulation and world models to reduce the gap.

Operators are increasingly locating work in lower-cost markets. Scale has opened a data factory in Mexico City, and Xdof has facilities in Shenzhen and Jakarta. Glassdoor data and job listings indicate robotics operators are paid about $25 an hour in the U.S., roughly double the level in Mexico, and Xdof chief executive Philipp Wu says the economics of large-scale data generation are difficult to justify in the U.S. alone.

That expansion is also prompting scrutiny over how footage is gathered and whether third-party data is reliable. Some factories in places such as Gurugram, India, are being used to collect video from workers wearing recording equipment, sometimes without extra pay, while some robotics executives say the most useful material is still collected internally. Researchers and start-ups continue to test whether larger datasets produce better model outputs, but Ken Goldberg of UC Berkeley says there is still no certainty that simply generating more data will deliver the hoped-for gains.

In our earlier article on TypeSafe AI’s Jev model and its $870 million funding round, we covered how the startup reached a $7.5 billion valuation just weeks after launch as enterprise adoption accelerated. We also explained that Jev focuses on automation beyond text generation, producing calibrated decisions in a transformer-based system designed to run faster and use fewer tokens than typical LLMs.

This material may contain third-party opinions, none of the data and information on this webpage constitutes investment advice according to our Disclaimer. While we adhere to strict Editorial Integrity, this post may contain references to products from our partners.