Macrodata Raises $4M Pre-Seed for Robotics Data Infrastructure

The critical need for high-quality data in advancing physical AI has spurred the launch of Macrodata, a new venture emerging from stealth with a $4 million pre-seed funding round. The company aims to build the foundational data infrastructure for the next generation of robotics, a sector experiencing explosive growth and investment. Leading this initial funding round was Air Street Capital, joined by a cohort of angel investors from prominent AI research institutions.

Macrodata is founded by Guilherme Penedo and Hynek Kydlíček, the architects behind Hugging Face's widely adopted FineWeb datasets. These datasets proved instrumental in the development of numerous open-source large language models, demonstrating the team's expertise in curating and refining massive data corpora. Their pivot to robotics signifies a recognition of the parallel challenges and opportunities in preparing real-world physical interaction data for AI training.

The robotics industry is witnessing unprecedented capital inflow, with venture funding reaching record highs in 2025 and projections for 2026 indicating even greater investment. Valuations for companies developing robotic hardware and AI control systems, such as Figure (valued around $39 billion) and Skild (valued around $14 billion), underscore the market's enthusiasm. This surge is fueled by advancements in vision-language-action models and world models, which are crucial for enabling robots to perceive, understand, and act within complex environments. However, scaling these sophisticated models hinges on access to vast, meticulously prepared datasets derived from real-world robot operations.

Unlike the relatively structured nature of text data, physical world data presents significant challenges. It involves large video files, disparate sensor sampling rates, interleaved action and language commands, and a lack of standardized formats. This fragmentation often forces development teams to create bespoke data processing pipelines for each new robot or sensor configuration, hindering rapid iteration and generalization. Macrodata's core offering, the open-source Python library Refiner, directly addresses this bottleneck.

Refiner is designed to ingest data from a variety of common robotics formats, including LeRobot, HDF5 (used in ALOHA, robomimic, LIBERO), Zarr, MCAP, raw video, and Hugging Face datasets. It transforms raw interaction logs into training-ready datasets by streamlining processes such as trimming unnecessary motion, annotating subtasks, tracking hand movements, and scoring trajectories using reward models and vision-language models. This capability allows for efficient data curation and enhancement, directly impacting the quality and performance of trained robotic policies. The platform offers a scalable cloud compute solution, enabling complex data processing tasks to be executed rapidly and cost-effectively.

The strategic approach of Macrodata mirrors the success of FineWeb: building an open-source core that establishes a de facto standard for robotics data management, complemented by a metered cloud service for large-scale processing. This model fosters widespread adoption within the developer community while creating a sustainable revenue stream. The team's proven track record in establishing data standards for LLMs provides strong confidence in their ability to replicate this success in the burgeoning field of physical AI, where robust data infrastructure is paramount for future innovation.