At NVIDIA GTC in San Jose, Milestone Systems has unveiled a major expansion of its Hafnia platform, introducing synthetic data capabilities designed to address one of artificial intelligence’s most persistent challenges: training models for the unknown.
The update enables developers to go beyond traditional datasets which are often limited to historical events by incorporating simulated scenarios such as rare weather conditions, unusual traffic patterns, and region-specific anomalies.
“AI systems typically learn from past events, but the real world is unpredictable,” said Edward Mauser, Director of Hafnia. “By combining trusted real-world data with synthetic augmentation, we can train models that are not only accurate, but resilient in unexpected situations.”
Bridging real and synthetic data
Hafnia, described as a bridge between data, training, and deployment, integrates curated real-world video with synthetic data generated through technologies such as NVIDIA Cosmos.
This approach allows developers to:
-
Simulate rare or dangerous events
-
Balance underrepresented data classes
-
Model regional and environmental variations
-
Reduce dataset bias
Crucially, Milestone emphasises that synthetic data is not a replacement for real-world datasets, but an enhancement layer that expands coverage while maintaining compliance and annotation quality.
Scaling AI development with NVIDIA
The expansion builds on collaboration with NVIDIA, leveraging its Physical AI Data Factory architecture a framework designed to unify data curation, augmentation, and evaluation at scale.
Using tools like NVIDIA Cosmos and OSMO, developers can transform raw data into high-fidelity, physics-aware training datasets, accelerating model development and improving reliability.
The result is a shift toward what Milestone describes as “proactive AI” systems capable of anticipating and handling edge cases rather than simply reacting to known patterns.
A focus on smart cities
Hafnia’s primary use cases centre on smart city environments, where unpredictability is the norm. From traffic management to public safety, AI systems must perform reliably across a wide range of conditions.
By expanding scenario coverage through synthetic data, Milestone aims to improve the robustness of computer vision models deployed in these environments.
Toward the next phase of AI training
The announcement signals a broader trend in the AI industry: moving beyond static datasets toward dynamic, continuously evolving training pipelines.
As organisations look to deploy AI in real-world, high-stakes environments, the ability to simulate the unexpected may prove just as important as learning from the past.
