Automotive manufacturers and research institutions need to amass millions of virtual miles for testing of AI based autonomous vehicles driving technology. However, research in autonomous driving in actual environment is hindered by infrastructure costs and the logistical difficulties of training and testing systems in a physical environment. An alternative option is to train and validate driving strategies in a simulated environment. Simulations can facilitate autonomous system testing for unforeseen scenarios and edge conditions in an efficient and cost-effective way.
We at ZATNav (Pvt) Ltd have embarked upon a project to create multiple environments (urban, forest, village and desert set ups) for the testing, training and validation of autonomous vehicles using latest machine learning and AI algorithms. This customizable and extendable setup will enable testing of autonomous vehicles using Unity3D Game Engine in combination with Unity Machine Learning Agents Toolkit / Python (Tensor Flow). Currently a prototype simulator for Autonomous Vehicle (AV) using AI algorithms for navigation has been prepared. The model of autonomous 3D vehicle learns to drive on its own using deep reinforcement learning. After several rounds of training using CNN, the vehicle can drive autonomously to complete the circuit.
A synthetic scenario has been created in unity 3D along-with a model of a car and a camera has been placed in front of the vehicle. It is apprized that Convolutional Neural Networks (CNNs) can be used for various computer vision tasks, including self-driving cars. In combination with socket programming for Unity 3D, we have developed a prototype system that receives data from Unity 3D. This data is processed using a CNN in Python, and sends back control commands to control the self-driving car. It is apprized that the current effort is just a proof-of-concept prototype, however, the goal is to develop a commercial Autonomous Vehicle Simulation Software (AVSS) that can generate realistic scenarios with ultra-high fidelity sensor simulations, providing realistic driving scenarios, and diverse environments. Once completed our simulation platform will facilitate testing, validation and quick time to market of AV using latest AI/ML algorithms by supporting the entire development process.
At ZATNav we have a Unity and 3D modelling team available, however some team augmentation will be required and for this we are actively seeking partners / investors to continue and the complete the project. The high-level project plan is as follows:
