Tech news in 3 minutes
Collecting robot training data is dirty, unglamorous work. Some AI labs are already paying XDOF to do it.
Two weeks ago, OpenAI announced it would relaunch its robotics program, shuttered in 2021, signaling a race among AI labs to teach machines to operate in the physical world. However, building capable robots requires training data that captures physical interaction—data that barely exists, unlike the vast text used for language models. YouTube videos and gig-worker footage are low-fidelity and hard to reconcile with the physical world. Enter XDOF (pronounced “ecks-doff”), emerging from stealth today. The startup aims to solve the data bottleneck for robotics by building data pipelines, collection tools, and annotation systems. It has raised $70 million from Thrive Capital, Spark Capital, a16z, Lux, and WndrCo. Co-founder and CEO Philipp Wu, a UC Berkeley PhD, previously worked on GELLO, a low-cost teleoperation system for generating training data. XDOF, founded in October 2024 with CTO Fred Shentu and COO Nemo Jin, has about 60 employees and 20 customers, including frontier AI labs. The company is releasing ABC, the largest collection of high-quality robot training data, with 130,000 trajectories, 300 hours of simulation, and 100 hours of evaluations. It plans to work across three tiers: teleoperation data on actual robots, general teleoperation data, and egocentric data from humans using wearable sensors. XDOF will hire and train teleoperators globally, arguing that major labs lack the focus and operational scale to do this themselves. The name XDOF stands for “unlimited degrees of freedom,” reflecting its ambition in physical AI.