With GenSim-2, developers can modify weather and lighting conditions such as rain, fog, snow, glare, and time of day or night in video data. | Source: Helm.ai
Helm.ai last week introduced the Helm.ai Driver, a real-time deep neural network, or DNN, transformer-based path-prediction system for highway and urban Level 4 autonomous driving. The company demonstrated the model’s capabilities in a closed-loop environment using its proprietary GenSim-2 generative AI foundation model to re-render realistic sensor data in simulation.
“We’re excited to showcase real-time path prediction for urban driving with Helm.ai Driver, based on our proprietary transformer DNN architecture that requires only vision-based perception as input,” stated Vladislav Voroninski, Helm.ai’s CEO and founder. “By training on real-world data, we developed an advanced path-prediction system which mimics the sophisticated behaviors of human drivers, learning end to end without any explicitly defined rules.”
“Importantly, our urban path prediction for [SAE] L2 through L4 is compatible with our production-grade, surround-view vision perception stack,” he continued. “By further validating Helm.ai Driver in a closed-loop simulator, and combining with our generative AI-based sensor simulation, we’re enabling safer and more scalable development of autonomous driving systems.”
Founded in 2016, Helm.ai develops artificial intelligence software for advanced driver-assist systems (ADAS), autonomous vehicles, and robotics. The company offers full-stack, real-time AI systems, including end-to-end autonomous systems, plus development and validation tools powered by its Deep Teaching methodology and generative AI.
Redwood City, Calif.-based Helm.ai collaborates with global automakers on production-bound projects. In December, it unveiled GenSim-2, its generative AI model for creating and modifying video data for autonomous driving.
Helm.ai Driver learns in real time
Helm.ai said its new model predicts the path of a self-driving vehicle in real time using only camera-based perception—no HD maps, lidar, or additional sensors required. It takes the output of Helm.ai’s production-grade perception stack as input, making it directly compatible with highly validated software. This modular architecture enables efficient validation and greater interpretability, said the company
Trained on large-scale, real-world data using Helm.ai’s proprietary Deep Teaching methodology, the path-prediction model exhibits robust, human driver-like behaviors in complex urban driving scenarios, the company claimed. This includes handling intersections, turns, obstacle avoidance, passing maneuvers, and response to vehicle cut-ins. These are emergent behaviors from end-to-end learning, not explicitly programmed or tuned into the system, Helm.ai noted.
To demonstrate the model’s path-prediction capabilities in a realistic, dynamic environment, Helm.ai deployed it in a closed-loop simulation using the open-source CARLA platform (see video above). In this setting, Helm.ai Driver continuously responded to its environment, just like driving in the real world.
In addition, Helm.ai said GenSim-2 re-rendered the simulated scenes to produce realistic camera outputs that closely resemble real-world visuals.
Helm.ai said its foundation models for path prediction and generative sensor simulation “are key building blocks of its AI-first approach to autonomous driving. The company plans to continue delivering models that generalize across vehicle platforms, geographies, and driving conditions.
Tell Us What You Think!