XDOF Exits Stealth Mode and Enters $1.2 Billion Series B Negotiations Within Three Months

A Robotics Data Startup's Rocket-Speed Valuation Growth

Robotics data startup XDOF is in negotiations for a Series B round at a $1.2 billion valuation — just three months after emerging from stealth mode. This news marks a new peak in market enthusiasm for data infrastructure in the robotics sector.

"Stealth mode" refers to the phase in which a startup operates secretly — developing technology and accumulating resources before publicly revealing its product and business model. Companies in stealth mode typically avoid media interviews and keep funding information private to prevent competitors from learning about their technical direction prematurely. "Unicorn" refers to a privately held company valued at over $1 billion, a term coined by venture capitalist Aileen Lee in 2013, when such companies were exceedingly rare. The fact that XDOF is approaching this threshold just three months after leaving stealth is historically remarkable — by comparison, even during the peak capital abundance of 2021, most AI startups needed 12 to 18 months to go from seed round to unicorn valuation.

XDOF's funding velocity — from stealth to unicorn-level valuation — stands out as exceptionally aggressive in the current AI and robotics landscape. Behind this rapid valuation inflation lies intense investor demand for robotics training data, a critical piece of foundational infrastructure.

The Strategic Value of the Robotics Data Track

XDOF focuses on the robotics data space, which is one of the core bottlenecks in today's embodied intelligence development. Embodied Intelligence refers to the ability of AI systems to interact with and learn from the real world through a physical body, representing one of the frontier research areas in artificial intelligence. Unlike AI in purely digital environments, embodied intelligence must handle real-world physics — gravity, friction, collisions — as well as multiple sensory modalities including vision, touch, and force perception. This makes its training data requirements far more demanding than those of traditional AI: an effective piece of robotic manipulation data must record not only joint angles and end-effector pose trajectories, but also synchronized force/torque sensor readings, depth camera point clouds, tactile sensor array outputs, and more — all with millisecond-level temporal alignment across data streams. As large language model technology matures, the intelligence bottleneck for robots has shifted from algorithms to high-quality, large-scale training data. The core challenge facing the industry today is the extremely high cost of data acquisition — one hour of high-quality manipulation data collected via human teleoperation can cost hundreds of dollars in equipment and labor, while training a general-purpose manipulation policy may require millions of trajectories.

Robotics data differs fundamentally from traditional image or text data. It must contain precise physical interaction information, multimodal sensor data, and temporal action sequences. Multimodal sensor data refers to heterogeneous information streams from different types of sensors — in robotics scenarios, typical modalities include RGB camera images, 3D point clouds from LiDAR or structured-light depth sensors, acceleration and angular velocity data from inertial measurement units (IMUs), position and velocity feedback from joint encoders, and increasingly important tactile sensing data (such as output from visuo-tactile sensors like GelSight). Temporal action sequences refer to the time-ordered control commands and state transitions recorded during task execution, forming the core training material for imitation learning and reinforcement learning. Managing this type of data presents unique challenges: different sensors operate at vastly different sampling frequencies (cameras typically at 30Hz, force sensors up to 1000Hz, and tactile sensors up to several thousand Hz), requiring specialized data alignment and compression solutions. The collection, annotation, and management of such data demand professional infrastructure — precisely the domain XDOF has chosen to enter.

From a market timing perspective, major players like Google DeepMind, Tesla, and Figure AI are all accelerating their humanoid robot and embodied intelligence initiatives, driving exponential growth in demand for high-quality robotics data. Specifically, Google DeepMind has released vision-language-action (VLA) models such as RT-2 (Robotic Transformer 2), which directly transfer large language model reasoning capabilities to robot control; its open-source Open X-Embodiment dataset consolidates over 1 million real manipulation trajectories from 22 robot platforms, making it one of the largest robotic manipulation datasets to date. Tesla's Optimus humanoid robot project is leveraging the visual neural network expertise accumulated through its autonomous driving program while deploying real robots in factories for data collection. Figure AI, having received joint investment from Microsoft, NVIDIA, OpenAI, and other tech giants, has seen its valuation exceed $2.6 billion, with its Figure 02 robot already entering commercial deployment. Additionally, companies like 1X Technologies, Apptronik, and Sanctuary AI are also accelerating their efforts. This competitive landscape has triggered explosive growth in demand for upstream data infrastructure. The fact that XDOF achieved such a high valuation in such a short time suggests that its technical approach or data assets have already been validated by leading industry clients.

Industry Signals Behind the Funding Pace

Going from stealth to Series B in three months is extremely rare. This typically implies one of several possibilities: the company completed substantial technical development and customer validation while in stealth; investor FOMO on the track has reached a peak; or the company possesses some scarce data resource or technological advantage. FOMO (Fear of Missing Out) is a powerful force in venture capital — when investors see a technology track that could define the landscape for the next decade, they often relax valuation standards to secure allocation. Since 2023, AI has seen multiple FOMO-driven mega-rounds: Anthropic has raised over $7 billion cumulatively, and xAI raised over $6 billion within just a few months.

This case also reflects a new trend in AI infrastructure investment: compared to competition at the model layer, data-layer infrastructure is emerging as the new investment hotspot. The underlying logic is clear: as model architectures converge (the Transformer architecture has become the de facto standard), data quality and data scale have become the core differentiating factors in AI system performance. Whether it's OpenAI's acquisition of Rockset to enhance its data retrieval and processing capabilities, or Databricks' $1.3 billion acquisition of MosaicML to vertically integrate its data platform with model training capabilities, both demonstrate the strategic importance of data infrastructure in the AI value chain. These acquisitions point in the same direction: in AI's "data flywheel," whoever controls the production and management infrastructure for high-quality data holds the long-term competitive advantage.

For the robotics industry as a whole, XDOF's rapid funding is a positive signal — it means the capital markets have recognized the commercial value of the robotics data niche. This will attract more resources into building infrastructure for embodied intelligence, accelerating the development of the entire industry.