CARLA+: An evolution of the CARLA simulator for complex environment using a probabilistic graphical model. Drones2023,7, 111. Instead of manually taking care of each condition, CARLA+ enables the user to automate the modeling of various environmental dynamics. Solving these challenges requires a newer goal of designing the evolved version of shared autonomous driving.
However, one of the challenging questions is whether autonomous vehicles are able to handle unexpected and sophisticated scenarios. Therefore, it is imperative that the AV has an enhanced perception of the environment by cooperating with neighboring vehicles using vehicle-to-. In this scenario, local perception of the environment is not sufficient for the AV to navigate safely and timely in these congested settings.
Therefore, it asks for dynamic adaptation of cooperation with other road users, including other vehicles. We believe that the previously mentioned scenes better capture the dynamics of the complex urban environment. There is always a trade-off between the accuracy of the 3D environment and the vehicle dynamics [9].
The proposed algorithms achieved significantly better results than the state-of-the-art algorithms.
Proposed Extension to CARLA
As the model size increases, the BN model derived from the set of examples becomes impractical. In this research, we chose the Hill Climbing (HC) search algorithm to learn the structure of BN. It is an abstract layer used to create different actors (vehicles, pedestrians, etc.), change weather conditions, get the current state of the world, etc.
In CARLA, a map consists of both the 3D model of the city and the road definition. Layered maps allow the user to switch between certain layers of the maps such as buildings, etc., while non-layered maps do not allow switching layers. Additionally, new user-defined maps can be created and imported into CARLA, allowing for greater customization and expansion of the system.
This would preserve the extensibility of the CARLA environment while providing the ability to model the environment more realistically and dynamically regardless of the map. Some of the parameters that can be set are cloud cover, precipitation, wind intensity, fog, sun azimuth, elevation angles, etc. Configuration files are an integral part of the framework, as they connect all other modules together and allow them to function smoothly.
Some of the predefined weather configuration files for some common scenarios, such as clear, rainy, cloudy, etc., are also provided. All of these configuration files are configurable and can be extended by the researchers, allowing more granular control over every aspect of the simulation environment. Each of these parameters can be configured via configuration files and can later be modified by the PGM model based on the state of the world.
This component was developed to capture the dynamics of pedestrians in the simulation environment. This is the main component of the CARLA+ that glues all other parts together. It takes the current state of the world as input and predicts the state of various variables based on it.
The simulated environment is then updated based on the most likely state of the desired variables. Therefore, instead of manually accounting for each state, CARLA+ enables the user to automate the modeling of different dynamics in the environment.
Validation of the Proposed CARLA+
As discussed in Section 4.3, the proposed CARLA+ is designed and developed to be used in a variety of ways to model dynamics in the environment. The purpose of this experiment is to model the dynamics between rain, the number of vehicles and their dynamic speed in the environment. More precisely, a model has been developed that can dynamically change the number of traffic and the speed of vehicles depending on the presence of rain in the environment.
The PGM model then predicts the traffic condition and speed based on the amount of rain in the simulation environment. These values vary based on the dynamic simulation environment and experiment settings and are easily configured via configuration files. Simple rules and dependencies are used to model the dynamic environment, which then generates values from random distributions to better model the randomness inherent in the real world.
Given that we are using a Poisson distribution instead of a hard-coded one to select the actual value, we ran the experiment 1200 times for each of the scenarios, i.e. NO_RAINandRAIN to better demonstrate the effectiveness of this approach. For this purpose, various features are collected from the real world and a BN is trained based on this data. The trained network is then used in the CARLA+ framework to simulate a dynamic real-world environment.
To prove the full potential of CARLA+, in this experiment we used real-world data to model the simulation environment. Traffic speed: we used New York City real-time traffic speed data [22] which consists of the traffic speed as well as the area where the data was captured. Structure learning is used to learn the structure of the BN network and to highlight the dependencies of different variables on each other.
The BN structure shown in Figure 10 shows that only two of the seven random variables, Time and Clouds, are independent, while all others are dependent on one or the other. Some of the parameters used to learn the structure and parameter learning are shown in Table 8. This trained model is then used to model real-world dynamics in the simulated environment allowing us to design dynamic and more realistic customized scenarios .
Because of the inherent randomness caused by the use of the Poisson distribution, we collected 100 samples regarding each transition of the different states of the variables, giving us a total of 51,200 data samples. These samples are then used to visualize and present the results of the trained model.
Conclusions
The open source code for CARLA+ is available at https://github.com/aadimator/CARLA-Plus (accessed 2 January 2023). PGM Probabilistic Graphical Model ROS Robot Operating System SAE Society of Automotive Engineers V2I Vehicle-to-Infrastructure V2V Vehicle-to-Vehicle V2X Vehicle-to-Everything References. In Proceedings of the Tith Annual ACM/IEEE International Conference on Human-Robot Interaction Extended Abstracts, Portland, OR, USA, 2-5. March 2015; pp.
Train here, drive there: Simulating real-world use cases with fully autonomous driving architecture in the Carla simulator. Train here, drive there: Validation of the ROS-based autonomous end-to-end driving pipeline in the CARLA simulator using the NHTSA typology.
