Milestone boosts Hafnia with synthetic data, TaaS & VLMs

Milestone Systems has expanded its Hafnia AI developer tools with synthetic data generation and a forthcoming Training-as-a-Service option, stepping up efforts to improve computer vision model training for rare and previously unseen situations.

The updates are being presented at NVIDIA GTC in San Jose, where Milestone is also outlining deeper integration between Hafnia and NVIDIA's data factory tooling for physical AI.

Hafnia is positioned as a bridge between curated data, model training, and deployment for vision AI projects, with a focus on smart city use cases. Milestone markets the platform as a way to reduce dataset bias and improve model performance across different environments and operating conditions.

Synthetic data

Synthetic data has become a key tool for developers who need broader coverage of edge cases than real-world datasets typically provide. City environments can include rare weather events, unusual traffic flows, region-specific vehicle types, and other scenarios that may not appear often enough in video archives to train models reliably.

Milestone is adding synthetic data to Hafnia's curated, real-world video library. The approach combines real footage with synthetic augmentation rather than replacing the real-world foundation.

"Together with NVIDIA, we are taking Hafnia to the next level by combining trusted real-world data with synthetic augmentation," said Edward Mauser, Director of Hafnia at Milestone Systems.

"This enables developers to train AI models that are not only accurate in known situations, but also resilient in the unexpected," Mauser said.

Milestone is using NVIDIA Cosmos Transfer for synthetic augmentation and says the goal is to expand coverage of rare or dangerous situations, improve representation of underrepresented object classes, model regional and environmental variation, and reduce dataset bias more systematically.

Milestone also emphasised compliance and traceability, saying Hafnia's synthetic data builds on a real-world dataset foundation with consistent annotation quality.

Training service

Alongside synthetic data, Milestone is previewing a Training-as-a-Service product for computer vision developers. It is positioning the service as a way to connect Hafnia's video library directly to training infrastructure.

Developers often assemble model training workflows from multiple components, which can raise costs and extend timelines. Milestone says its approach provides access to Hafnia's compliant, traceable video library, with both real-world and synthetic datasets available for use in customers' own pipelines.

Users will be able to customise datasets and fine-tune models for specific use cases, Milestone said, adding that traceability of data sources can help teams align with relevant regulations.

Milestone says Hafnia can reduce time spent sourcing and managing training data, leading to analytics solutions being built up to 30 times faster. It did not provide details on how it measured the claim.

VLM platform

Milestone also outlined its VLM-as-a-Service offering, built in partnership with NVIDIA. It provides hosted visual language models based on NVIDIA Cosmos Reason models and tuned for smart city environments, according to Milestone.

At GTC, Milestone is announcing the availability of a new EU-optimised visual language model for traffic applications. The model is already running with selected EU city customers, and Milestone says more models will follow for additional smart city scenarios.

Milestone says hosted VLMs reduce the need for repeated retraining and lower costs tied to data collection and infrastructure growth. It also reported performance improvements over base models, including a 19.4% improvement in flow and direction correctness, an 8.9% improvement in visual feature detection, and a 4.4% improvement in alert verification accuracy.

Milestone also said teams can accelerate the final stages of the lifecycle by 70 times with its VLM-as-a-Service, but did not disclose the baseline used for comparison.

Cloud approach

Hafnia runs on a multi-cloud setup that includes AWS and Nebius, alongside other providers. Milestone says the strategy combines general cloud infrastructure with specialised AI computing resources, aiming to match compute capacity to different phases of model development.

The synthetic data generation pipeline-using Cosmos Transfer and Cosmos Evaluator powered by Cosmos Reason-is being launched on Nebius. Milestone also cited data sovereignty as a factor in its cloud design, saying customers can control where sensitive information is stored and processed.

Milestone says its partnerships and infrastructure cover the full lifecycle of vision model development, from data sourcing and custom fine-tuning to training and deployment.

Mauser is due to present a session titled "A New Frontier for Vision AI with Expert Reasoning Agents" at the event.