microagi Ships Petabyte-Scale Robotics Data Pipe at 10 GB/s

Munich startup microagi has launched a delivery system for its human-demonstration datasets, with direct writes to AWS, Google Cloud, Cloudflare R2, and the Hugging Face Hub. The company says the system can sustain more than 10 gigabytes per second for each customer and move as much as one petabyte per day. For teams training robot foundation models, which are broad models adapted to many physical tasks, the service targets the costly delay between selecting data and making it available to a training cluster.

Customers filter recordings through a web portal and choose a destination. microagi says the first bytes arrive within minutes. The company worked with Hugging Face and Cloudflare on the transfer layer, which supports several customer deliveries concurrently.

A corpus that strains ordinary delivery

microagi records people performing manual work, packages those demonstrations as training data, and helps customers adapt robotics systems to specific factory tasks. Its contributor network includes 20,000 people across 15 countries, according to the company.

Each recording can combine standard color video, depth measurements, pose estimates, and inertial measurement unit data from sensor-equipped gloves. IMUs capture motion and orientation, while pose data tracks the position of hands or tools. A corpus containing millions of hours can produce customer selections measured in hundreds of terabytes.

Transfers at that scale create operational problems beyond raw storage capacity. Copying data through an intermediate bucket adds another full transfer, consumes temporary storage, and delays preprocessing. Direct delivery removes that staging step and places the selected files in the cloud environment where training will run.

From portal to training bucket

The new platform adds a self-service control plane and a high-throughput transfer path to microagi’s existing catalog. Its main capabilities are:

• Corpus filtering: Customers select relevant tasks and recordings through a web interface.
• Fast startup: microagi says delivery begins within minutes of a request.
• Per-customer throughput: The company claims more than 10 GB/s for each active customer.
• Concurrent delivery: Several customers can receive data at the stated rate at the same time.
• Direct destinations: Supported targets include AWS S3, Google Cloud Storage, Cloudflare R2, and the Hugging Face Hub.

The Hugging Face integration places the data alongside tools that robotics teams may already use, including LeRobot, an open library for training robot policies from demonstrations. Cloudflare supplies infrastructure for distributing the data across destinations.

Using decimal units and excluding protocol overhead, the published figures imply the following transfer capacity:

These figures are company claims rather than independent benchmarks. Actual completion times will vary with file counts, destination limits, network conditions, retries, and validation work.

Robotics needs purpose-built data

Robot-learning teams cannot draw on a web-scale archive of physical manipulation in the way language-model developers can collect text. Useful demonstrations require cameras, calibrated sensors, consistent task definitions, and people performing actions in relevant environments. Companies including Scale AI, Turing, and Micro1 also sell human-generated training data, with different levels of robotics specialization.

NVIDIA has expanded the available supply through its Human 1M dataset, which contains synchronized video and 3D hand tracking. Its open robot-learning workflows for DGX Station also list microagi among the supporting solution providers. These efforts reflect demand for demonstrations that capture real motion, contact, tools, and environmental variation.

microagi combines data collection with on-site deployment support. That approach gives the company a role in adapting an existing robot and model to a customer’s production process, while many data vendors focus primarily on sourcing and annotation.

Fresh capital meets factory scale

microagi recently closed a $55 million seed round led by Hummingbird, with participation from Northzone, LocalGlobe, Village Global, and redalpine. The company describes it as the largest seed round raised by a German startup.

Five companies are collecting data through the platform, according to microagi, and one is preparing to deploy robots on a production line. Its customers operate in automotive manufacturing, logistics, and food production. Moving from pilot recordings to production-scale training increases both the volume of demonstrations and the frequency of dataset updates, which makes delivery infrastructure part of the training workflow.

The last mile into training

A visuomotor policy maps camera and sensor inputs to robot actions, while a world model learns how an environment may change after an action. Both approaches can consume large collections of synchronized recordings. Faster delivery shortens the interval between choosing a dataset and beginning validation, preprocessing, or training.

Teams using Hugging Face may be able to connect a delivered repository to datasets.load_dataset(), subject to the dataset configuration, access controls, and file format. AWS and Google Cloud users can point existing ingestion jobs at their own buckets after registering manifests and validating the delivered objects.

Direct delivery reduces copying, but training readiness still depends on checksums, schemas, sharding, indexes, permissions, and preprocessing. Cloud placement also affects cost. Writing into a customer’s preferred provider can reduce duplicated storage, while source egress, cross-region traffic, and destination request charges may still apply.

Before the first training run

microagi’s public material provides headline throughput and destination support, but production buyers will need implementation details in several areas:

• Benchmark conditions: Average object size, file count, regions, transfer duration, and sustained versus peak throughput.
• Integrity controls: Checksums, retry behavior, partial-transfer recovery, and completion manifests.
• Security: Encryption, credential handling, tenant isolation, audit logs, and customer-managed keys.
• Dataset lifecycle: Versioning, incremental updates, deletion requests, and reproducible snapshots.
• Cloud costs: Egress, storage, API requests, and any platform fees attached to delivery.
• Data quality: Contributor consent, licensing, sensor calibration, task labels, and provenance metadata.

microagi collects demonstrations, delivers them, and supports deployment integration. Customers and partners supply the robot hardware and train the foundation models. The new transfer layer can shorten data ingestion if its published rates hold under realistic object counts, while reliable metadata, predictable costs, and verifiable integrity will determine how smoothly those bytes enter a production training pipeline.