Accelerating the data flywheel for physical AI: How NormaCore is standardizing robot fleet management

Fleet-scale learning refers to the process of using every hour of robot operation to improve a shared model. It requires two things that the robotics industry does not yet have: a runtime that unifies hardware, and a storage layer that can keep telemetry at full fidelity without bankrupting the operator.

Today, researchers spend most of their time on infrastructure instead of science. Most telemetry is discarded. Training pipelines are rebuilt from scratch for every new hardware platform. The result is that the data flywheel for physical AI is not optimized when compared to that of generative AI, where more data produces better models which attract more users, and generate more data.

NormaCore is building the infrastructure to accelerate the flywheel by orders of magnitude.

“Robotics has no standard data layer. Every manufacturer ships a different SDK. Every lab hand-builds its own pipeline. Most telemetry is thrown away because storing it in full fidelity is too expensive. As a result, researchers spend their time on infrastructure instead of science, and the industry cannot build the data flywheel that physical AI needs”

— Alex Batsuev, Co-Founder, NormaCore

The architecture of NormaCore is rooted in a fundamental insight: the next frontier of computing is the physical world. Recognizing this early on, Alex Batsuev and Dmitriy Samsonov applied the rigor of backend systems development to robotics infrastructure. Starting in early 2022, they focused on creating “zero-cost” abstractions — highly efficient systems that hide complexity without sacrificing performance. Through a series of technical innovations, NormaCore now powers vision-language-action (VLA) models, offering a universal, fragmentation-free environment for physical AI operations.

NormaCore has three main areas of innovation: a unified runtime, a custom database engine, and a 3D-printable reference arm. Station is an open-source runtime that unifies fragmented hardware SDKs into a single, consistent interface. Station blurs the line between research and production: developers use the same tools for rapid prototyping and fleet deployment, removing the costly rewrite that typically separates the two phases.

The NormaCore stack is designed as a closed loop. Station, running on the robot, captures demonstrations and operational telemetry directly into NormFS. NormFS compresses and encrypts that data locally, then synchronizes it to Nebius object storage over a standard S3-compatible API. When a dataset is ready for training, NormaCore deploys serverless GPU instances on Nebius, runs the training job, and pulls the resulting policy back to the fleet, all through a single provider.

NormaCore needs two things from a cloud provider: affordable, reliable object storage for the telemetry that NormFS captures, and on-demand GPU compute for training VLA models that turn that data into deployable policies.

Most general-purpose clouds price storage and compute separately in a way that penalizes the continuous, high-volume workloads that robotics generates. Telemetry doesn’t arrive in neat batches. Instead, it streams constantly from every sensor on every robot in a fleet. In addition, training runs don’t follow a predictable schedule; they spin up when enough new data warrants an update. Infrastructure designed for web workloads is poorly suited to this pattern.

Nebius provides both through a single integration: S3-compatible storage for long-term telemetry retention, and serverless GPU compute for training and inference. For NormaCore and its users, this means one provider, one bill, one integration point, and a unified workflow from data capture to model deployment. Streamlined integration helped NormaCore transition from initial evaluation to full production use in less than 3 hours.

“A researcher running Station on an SO-101 arm can collect demonstrations into NormFS, push the dataset to Nebius object storage, spin up GPU instances for training, and pull the resulting policy back to the robot without stitching together three different vendors”

— Dmitriy Samsonov, Co-Founder, NormaCore

The Nebius platform also offered the vital capability to store and process high-fidelity telemetry data strictly within Europe, a secure and compliant foundation for an expanding global user base. The single-provider model has shifted the team’s day-to-day focus from engineering plumbing to advanced scientific exploration of into research workflows that were previously impossible. For example, the team has furthered its investigation into distributed learning capabilities with EGGROLL (Evolution Guided General Optimization via Low-rank Learning, an algorithm for training massive AI models).

NormaCore’s broader ambition is to develop a foundation model that improves with every robot in every fleet and uses the most diverse, highest-fidelity telemetry dataset in the industry. The infrastructure to make that possible is being built now, on an open-source runtime, a custom data engine, and a cloud that doesn’t penalize the workloads that physical AI demands.

NormaCore is recruiting commercial hardware partners for integration, with the goal of expanding the platform from the open-source community to production fleet deployments. Over the next 6–12 months, NormaCore will launch experimental deployments in key industries such as agriculture, logistics, and retail.