Helm.ai has expanded its autonomous driving software platform, introducing a vision-based system designed to scale from Level 2+ driver assistance to Level 4 urban autonomy without relying on lidar sensors or high-definition maps.
The updated platform, known as Helm.ai Driver, is built on the company’s Factored Embodied AI architecture. The system uses camera-based perception and machine learning to enable automated driving capabilities in complex urban environments.
According to Helm.ai, the software is designed to allow automakers to deploy Level 2+ driver assistance features today and later expand toward Level 3 and Level 4 autonomy using the same core architecture as hardware capabilities and regulatory frameworks evolve.
The company released a demonstration video showing the system navigating city streets in Redwood City, California. The test vehicle performed turns at intersections, followed traffic signals and interacted with surrounding vehicles while supervised by a safety driver.
Vladislav Voroninski said the autonomous driving industry is facing increasing challenges related to data requirements for improving system performance.
“The industry has reached a tipping point,” Voroninski said, adding that large-scale data collection alone is no longer economically viable for developing advanced autonomy.
Helm.ai’s system separates autonomous driving into two main software layers: a perception layer that interprets sensor data into semantic and 3D environmental information, and a policy layer that determines vehicle behavior based on that structured understanding of the road.
The company said this structure is designed to improve interpretability and transparency, an important factor for safety certification and regulatory approval for higher levels of autonomy.
Helm.ai also said its planning system achieved urban driving capability using about 1,000 hours of real-world driving data through a training approach it calls Deep Teaching™, which allows neural networks to learn from large-scale non-driving visual datasets.
The architecture is intended to support deployment across different regions without location-specific retraining. The company said it tested this capability by deploying the system in Torrance, California without prior local data training, demonstrating what it described as zero-shot generalization.
Helm.ai said the approach could allow automakers to scale autonomous driving features globally without relying on geofenced deployment areas or extensive city-by-city data collection.
