XDOF negotiates a $1.2 billion valuation as the robot-training race mirrors Scale’s real-world experiment

The startup XDOF is in advanced talks to close a second-round (Series B) financing led by 8VC, at a valuation of about $1.2 billion, less than three months after emerging from stealth mode. The negotiations are driven by rapid financial growth that has pushed the company’s annual revenue toward the $50 million threshold, prompting venture investors to request the round ahead of the schedule the company had planned, following a first round that raised $70 million with participation from funds including Thrive Capital, Andreessen Horowitz, Lux and Spark Capital.

Physical data supply chains become the backbone of building general-purpose robots.The company was founded in 2024 by researchers from the University of California, Berkeley, Philip Wu and Fred Shinto, with the aim of solving the major dilemma facing intelligent robot models. Unlike large language models, whose initial training relied on the flood of text published online, physical robots lack an equivalent repository of real-world data, turning the collection of motion and interaction data into the primary bottleneck to building multipurpose machines capable of operating in unprepared environments.

The core technology originated from Wu’s PhD research on machine learning from massive data, where he partnered with Shinto to develop the GELLO system, a low-cost remote operating system that lets a human operator control a robotic arm to generate high-quality training data. The company is now expanding to build pipelines and tools for collecting and labeling data for leading AI labs and robot manufacturers, acting as an external supply chain specialized in motion data, similar to the role played by companies such as Scale AI and Mircor in training language models, amid competition from other platforms seeking to enter the field such as Meka AI and Micro 1.

The company is partnering with the AI research lab at Berkeley to launch a data package called ABC, which it describes as the largest high-resolution dataset dedicated to training robots. The collection method relies on two integrated streams: the first uses operators who remotely guide robots, and the second employs teams of field data collectors who wear sensory probes on their bodies to record details of everyday motions and tasks, such as folding fabrics and flattening cardboard boxes, with plans to hire and train specialized human networks worldwide to gather this motion data and expand its client base, which currently includes twenty entities, among them major AI labs.

This shift shows development teams and companies in the Gulf, Egypt and the region that the real challenge in automating supply chains, warehouses and industrial processes no longer lies solely in inventing software architectures, but in providing realistic motion data tied to local work environments. As the data economy moves from classifying textual code to recording physical movements and directing robotic arms, investing in field-data collection infrastructure and attracting remote-control operators becomes the decisive path to lowering robot deployment costs and adapting them to daily tasks in regional markets.