Skip to main content

XDOF

Category: Robotics / Embodied AI

XDOF builds data pipelines, collection tools, and annotation systems for training robot foundation models, solving the physical-world data scarcity bottleneck in robotics AI. XDOF was founded in 2024. The company is led by Philipp Wu. Based in Berkeley, California, United States. Team size: 51-100. Total funding raised: $78.3M. Latest round: Series A. Key investors include Thrive Capital, Spark Capital, Andreessen Horowitz (a16z), Lux Capital, WndrCo.

Founded
2024
Headquarters
Berkeley, California, United States
Team size
51-100
Total funding
$78.3M

Value proposition

XDOF provides the physical-world training data that AI labs need to teach robots how to interact with their environment — solving the critical data bottleneck that prevents robotics foundation models from scaling. They offer a complete outsourced solution: data pipelines, collection tools, annotation systems, and policy training, so frontier labs don't need to build their own warehouse-scale teleoperation infrastructure.

Products and solutions

1) Teleoperation data pipelines (bespoke robot-specific data collection), 2) Generalized teleoperation data (GELLO-based), 3) Egocentric human demonstration data with proprietary wearable sensors, 4) Data cleaning and annotation services, 5) ABC-130K open-source bimanual manipulation dataset (130K+ trajectories, 300 hours simulation, 100 hours evaluations), 6) Policy training support

Unique value

XDOF is the only company building end-to-end data infrastructure specifically for robot foundation models — combining proprietary hardware (wearable sensors, teleoperation rigs), large-scale data collection operations, annotation pipelines, and policy training into a single outsourced solution for frontier AI labs.

Target customer

Frontier AI labs, robotics companies, and research institutions building robot foundation models that need large-scale, high-quality physical manipulation training data.

Industries served

Robotics, Artificial Intelligence, Physical AI, Automation

Technology advantage

Rooted in GELLO teleoperation framework (influential UC Berkeley research paper); full-stack capability from hardware (wearable sensors, teleoperation rigs) through data ops to policy training; proprietary hand-tracking algorithms; three-tier data pyramid strategy; largest open-source bimanual manipulation dataset (ABC-130K); team includes alumni from Covariant, Meta, and Tesla

How they differentiate

Unlike competitors building robot foundation models or hardware, XDOF focuses exclusively on the data infrastructure layer — building end-to-end pipelines from hardware (proprietary wearable sensors, GELLO teleoperation rigs) through data collection, cleaning, annotation, and policy training. They released ABC-130K, the world's largest open-source bimanual robot manipulation dataset (130K+ trajectories), as a showcase of their data capabilities. Their full-stack approach spans hardware design, operations, and policy training.

Main competitors

Physical Intelligence, Covariant, Field AI

Key partnerships

UC Berkeley AI Research Lab (ABC-130K dataset co-release), Carnegie Mellon University, MIT, Amazon FAR (Fulfillment Automation & Robotics)

Notable customers

~20 active customers including several frontier AI labs (names not publicly disclosed)

Major milestones

October 2024: Company founded, June 2026: Emerged from stealth with $70M funding, June 2026: Released ABC-130K, world's largest open-source bimanual robot manipulation dataset, June 2026: 20 active customers, 60+ employees

Growth metrics

~60 employees; ~20 active customers (including multiple frontier AI labs); launched with $70M funding in June 2026

Market positioning

XDOF is positioning as the outsourced data infrastructure layer for the robotics AI industry — building the specialized data pipelines, collection tools, and annotation systems that frontier AI labs and robotics companies need but cannot easily build themselves. It operates across a three-tier "data pyramid": bespoke teleoperation data, generalized teleoperation data, and egocentric human demonstration data.

Geographic focus

United States (primarily Bay Area / Berkeley); global teleoperator workforce planned

About Philipp Wu

PhD student at UC Berkeley advised by Prof. Pieter Abbeel (Robot Learning Lab). Research at intersection of reinforcement learning, unsupervised learning, and robotics. Co-authored GELLO teleoperation framework, DayDreamer (CoRL 2022), Masked Trajectory Models (ICML 2023). Previously worked on BLUE low-cost robotic manipulator project.

Official website: