Robotics Data Startup XDOF Secures $70 Million and Open-Sources Landmark Training Dataset

As the artificial intelligence industry races to build machines that can operate in the physical world, a critical bottleneck has emerged: the lack of high-quality training data. While large language models achieved breakthroughs by ingesting the entire internet, robots require precise, physical interaction data that simply does not exist online. To solve this "chicken-and-egg" problem, robotics data infrastructure startup XDOF officially emerged from stealth this week, announcing a $70 million funding round.[1][4]

The round drew participation from a roster of heavyweight venture capital firms, including Thrive Capital, Spark Capital, Andreessen Horowitz (a16z), Lux Capital, and WndrCo. Founded in October 2024 by Philipp Wu, Fred Shentu, and Nemo Jin—alumni of UC Berkeley, Tesla, and Meta—XDOF operates as an outsourced data factory for the robotics industry. The company builds the specialized data pipelines, collection hardware, and annotation systems required to train robots for real-world physical interaction.[2][3][4]

"We didn't have large-scale data to work with," CEO Philipp Wu explained, recalling his time as a Ph.D. student. "We first needed to actually collect data before we could even ask how to train a foundation model for robotics." To prove the efficacy of its infrastructure, XDOF simultaneously released ABC-130K, which it describes as the world's largest open-source bimanual robot manipulation dataset.[1][5]

Developed in collaboration with researchers from UC Berkeley, Carnegie Mellon University, MIT, and Amazon, the landmark dataset provides the academic community with an unprecedented foundation for training physical AI. The release includes 130,000 robot manipulation trajectories, representing 3,500 hours of real-world interaction data. The dataset spans 195 distinct physical tasks, capturing a wide spectrum of manipulation primitives such as pick-and-place actions, handovers, and tool use.[2][5][6]

More impressively, the dataset includes highly dexterous behaviors that have traditionally confounded robotic systems, such as folding T-shirts, flattening cardboard boxes, and precisely placing AirPods into their plastic charging cases. By open-sourcing this massive corpus, XDOF aims to establish a universal baseline, allowing researchers to iterate on model designs without spending millions of dollars to build their own data collection warehouses.[1][5][6]

Generating this caliber of data requires a massive, labor-intensive operation. Existing workarounds, such as scraping YouTube videos or using low-quality footage captured by gig workers, have proven inadequate because they lack the precise spatial fidelity and control inputs needed for effective robotic learning. Instead, XDOF employs a three-tier data acquisition strategy. The primary method involves direct teleoperation, where trained human operators use specialized rigs to remotely control robotic arms, effectively demonstrating the exact physical movements the AI needs to mimic.[1][2][4]

The company's secondary and tertiary data collection tiers involve lower-cost teleoperation devices—stemming from an open-source project called GELLO—and egocentric wearable sensors that capture everyday human movements from a first-person perspective. This layered approach allows XDOF to scale its data production efficiently, transforming a bespoke research experiment into a standardized, industrial-grade infrastructure business.[1][2]

The market demand for this infrastructure is already evident. Despite operating under the radar for nearly two years, XDOF has grown to 60 employees and secured approximately 20 active customers, including several of the world's leading frontier AI laboratories. The fact that these well-funded labs are choosing to pay XDOF rather than build their own internal pipelines reveals a deliberate strategic shift. By outsourcing data collection, AI developers can keep the operational complexity of maintaining warehouses, calibrating hundreds of robots, and managing teleoperators off their balance sheets.[1][2][4]

XDOF's launch arrives at a pivotal moment for the broader technology sector. Just weeks ago, OpenAI announced the revival of its internal robotics training program—an initiative it had previously shuttered in 2021 to focus exclusively on software models. This pivot signals a growing consensus that "physical AI" is the next major frontier. With its massive open-source contribution and robust commercial pipeline, XDOF is positioning itself as the foundational utility for the incoming robotics revolution.[1][4][5]