Testing Autonomous Driving Systems: Co-simulation and Human Driver Calibration
21st October, 2024
Alkis Papadoulis
Scientific Researcher, Aimsun
The journey to fully automated driving is an exciting one, with innovations making cars smarter, safer, and more efficient. The EU-funded research project, i4Driving, contributes to this journey by developing a cutting-edge platform designed to test Autonomous Driving Systems (ADS).
The Challenge: Testing Autonomous Driving Systems
Autonomous vehicles must navigate complex environments, make split-second decisions, and interact seamlessly with human drivers. However, one of the major challenges in testing ADS is representing human driving behavior accurately. Human drivers exhibit unpredictable reactions, unique decision-making processes, and varying levels of experience. As a result, creating a comprehensive testing environment for ADS requires a system that accurately mirrors the real world and its human factors.
This is where i4Driving comes into play.
Calibrating Human Driver Models for ADS Testing and sensitivity analysis
The team captured human driving behavior at real-world test tracks and driving simulators, including decision-making under different conditions. This contributed to a robust dataset that helps calibrate the new i4Driving human driver model that can form the new baseline for autonomous vehicle testing.
More specifically, the team is working on developing a fully automated calibration and validation pipeline that will use artificial intelligence algorithms to determine the optimal parameters that represent human drivers accurately in Aimsun Next. This pipeline, in a slightly different form, has been used already to perform sensitivity analysis.
Specific scenarios with initial conditions have been formulated and cutting-edge sensitivity analysis methodologies have been applied in order to understand the contributing factors to safety-critical scenarios for the project.
Below, you can see the scenario investigated a few months ago in order to determine critical conditions that would lead to a crash in a car following situation.
For this scenario, the leader decelerates to a full stop with varying deceleration rates. Another variable in this scenario is the initial speed of the two vehicles.
- THW in the figure stands for the Time Headway between the two vehicles that is constant in this scenario at 2 seconds.
- a is the minimum intervehicle time gap, and
- b the car-following time gap.
The results of this sensitivity analysis test can be seen above. The number of samples was set to 200 arbitrarily. The sampling method selected for this experiment was the Latin hypercube. They indicate that collisions (shown with black stars in the figure) occur in situations of high leader deceleration and high vehicle speed.
Co-Simulation with CARLA: Merging Virtual and Real-World Testing
Once the human driver model has been calibrated and validated using the tools above, the effectiveness of ADS needs to be evaluated in an integrated simulation environment. In other words, we need to place an actual Autonomous Vehicle in an environment of human driven vehicles with the i4Driving model in Aimsun Next. In order to do this, we developed an all new interface between CARLA and Aimsun Next.
CARLA is a high-fidelity open-source simulator for autonomous driving research. Our co-simulation platform allows for testing ADS in realistic, virtual environments that simulate various traffic, weather, and road conditions. In addition to that, the CARLA simulator allows all three of autonomous vehicle layers to be present in the co-simulation environment, allowing OEMs to effectively test their systems, while gaining the benefits of the sophisticated traffic environment of Aimsun Next. The combination of real-world data and simulation provides a comprehensive testing ground for autonomous systems.
CARLA’s advanced simulation capabilities enable us to include an ADS in the Aimsun Next environment, allowing ADS developers to test edge cases and complex situations in a controlled, safe environment. By leveraging CARLA, we can push the boundaries of ADS testing, offering a dynamic and scalable way to validate systems before they are deployed on public roads.
The image below shows the network developed for an initial test of the platform. This is one of the default maps included in CARLA, but also translated into Aimsun Next.
For this experiment, a simple autonomous vehicle controlled by CARLA was surrounded by traffic produced by the Aimsun Next transport simulation platform. You can see the two environments running side by side in the figure below. The autonomous vehicle in this case is the vehicle in front, with Aimsun Next generating the vehicles following closely behind.
We are aiming to polish and release the code to pair Aimsun Next and CARLA. Watch this space for more updates and demonstrations!
The post Testing Autonomous Driving Systems: Co-simulation and Human Driver Calibration appeared first on Aimsun.