Robotics Startup XDOF Emerges With $70M to Solve AI's Physical Data Bottleneck

The artificial intelligence industry has spent the last three years conquering digital tasks, but teaching a machine to fold a t-shirt remains a monumental challenge. Now, a newly launched startup called XDOF is emerging from stealth to solve the physical data bottleneck holding back the robotics revolution. Backed by a massive $70 million seed round, the company is building the outsourced data infrastructure required to train the next generation of general-purpose robots.[1][2]

The funding round, which officially closed this week, drew participation from a roster of heavyweight venture capital firms. Thrive Capital, Spark Capital, Andreessen Horowitz (a16z), Lux Capital, and WndrCo all backed the San Mateo-based startup. The sheer size of the seed investment underscores a growing consensus in Silicon Valley: the next defensible, highly lucrative layer of the AI boom will be deeply physical.[3]

Founded in October 2024 by UC Berkeley alumni Philipp Wu, Fred Shentu, and Nemo Jin, XDOF operates on a simple premise. While frontier AI laboratories excel at developing complex software models, they lack the operational scale and desire to independently manage massive, messy physical data operations.[2]

Physical manipulation data is the specific bottleneck preventing frontier AI labs from training generalist robots. Building an in-house data pipeline requires hundreds of thousands of square feet of warehouse space, fleets of expensive robotic hardware, continuous mechanical calibration, and a globally distributed workforce of trained human operators.

Rather than forcing software companies to become heavy-industrial operators, XDOF serves as an outsourced data factory. The company provides end-to-end data pipelines, collection hardware, and annotation systems, allowing AI labs to keep warehouse-scale operational complexity off their balance sheets while still advancing their embodied intelligence programs.

To demonstrate its capabilities and immediately impact the broader research community, XDOF has partnered with UC Berkeley's AI Research lab to release ABC-130K. The company describes the release as the largest and highest-quality open-source robot manipulation dataset ever made available to the public.[1][2]

The ABC-130K dataset contains 130,000 distinct physical trajectories, supplemented by 300 hours of simulation data and 100 hours of rigorous evaluation metrics. The dataset captures robots performing tasks that require extreme precision and spatial awareness, such as flattening cardboard boxes, folding laundry, and carefully placing wireless earbuds into their charging cases.[1]

Gathering this volume of high-fidelity physical data requires a multi-tiered acquisition strategy. XDOF deploys a three-pronged approach to capture human-level dexterity and translate it into machine-readable formats that foundational models can easily ingest.

The first tier involves direct teleoperation, where human operators remotely control the exact deployment robots that will eventually run the AI models. The second tier utilizes "GELLO" devices—specialized, low-cost teleoperation rigs that mimic robotic joints. The final tier relies on egocentric wearable sensors worn by humans as they go about everyday tasks, capturing the subtle mechanics of human movement.

The startup's infrastructure-as-a-service model is already proving highly attractive to the industry's biggest players. Despite operating under the radar for less than two years, XDOF has already secured approximately 20 active enterprise customers, including several leading frontier AI research groups.[1]

The timing of XDOF's launch aligns perfectly with a broader industry pivot toward physical AI. Just weeks ago, OpenAI announced the revival of its own robotics training program, signaling that the race to build embodied intelligence is accelerating. By standardizing data collection and cleaning, XDOF aims to eliminate the historical lag between rapid robot hardware advancements and the software required to operate them.[1]

With $70 million in fresh capital, XDOF plans to rapidly scale its operations. The company will use the funds to hire and train a global workforce of teleoperators and egocentric data gatherers. Additionally, XDOF is developing its own proprietary wearable sensors to ensure that its hand-tracking algorithms perfectly match the mechanical realities of the robots being trained.[1]