Advanced simulation for autonomous vehicles

Autonomous vehicles will be among the most complex mechatronic systems ever created. Advanced artificial intelligence (AI) and intelligent sensors, running on bleeding edge integrated circuits (ICs) and systems-on-chip (SoCs), will control complex mechanical and electrical systems as they guide the vehicle through dynamic driving environments based on inputs from a vast array of sensors. The initial design of these systems certainly will be an immense challenge, but verifying and validating these systems is expected to be even harder. A major challenge is the cross-domain nature of autonomous vehicles: mechanical, electrical, electronics, and software all have to work in tight harmony to achieve self-driving functionality. In addition, these various systems have to be given inputs on which to act. As a result, the verification and validation of these vehicles must adopt an equally integrated approach in which each domain is verified in the context of each of the others, and in a realistic driving environment.

The verification and validation needs of such vehicles far outstretch the capabilities of most simulation platforms available today. As a result, there is a need for closed-loop simulation platforms that not only provide the virtual environment required to develop the next generation of mechatronic systems, but also to develop the computing and software solutions that will maintain control.

Jim McGregor of TIRIAS Research recently published a whitepaper detailing such a simulation platform, titled AV Simulation Extends to Silicon. You can read the paper in full, here. The following is an excerpt from that whitepaper.

The challenge of automated driving

The dream of automated vehicles emerged in the 1950s and some of the first self-driving vehicles were developed in the 1980s, but it has only been within the past few years that a mass production vision solution for automated vehicles has taken shape. Advancements in Artificial Intelligence (AI), sensors, interconnects, and semiconductor technology combined with the electrification of vehicles have created the perfect formula for vehicle automation at all levels.

Even with these advancements, creating an automated vehicle is no small challenge. Estimates for training the AI neural network models for automated driving range from 10 billion to 20 billion miles of driving; this alone would require hundreds of years to accomplish in the real world. Even after the AI models are trained, the operation of the vehicle must be validated in a repeatable manner for all possible vehicle configurations, operating variables, and driving scenarios. Among these considerations are different road and weather conditions, different loads or the shifting of loads within the vehicle, tire and brake conditions due to temperature and wear, and faulty sensors, actuators, and electronics control units (ECUs), to name just a few. Accounting for all possible factors to ensure the safe operation of vehicle automation at any level seems an overwhelming task. However, it must be achieved to ensure the safety of the occupants, pedestrians, and others that may come within the vicinity of the vehicle.

For additional context, here I explain how virtual validation techniques can lead to correct decisions in autonomous vehicles.

The age of simulation

Just as simulators are used to train pilots and automated flight systems for a wide variety of possible flying conditions, the automotive industry is relying on simulation to overcome the challenges associated with the training, testing, and validation of automated vehicles. Through the simulation of environments, vehicles, and driving scenarios, neural network models can be trained and tested without the risk and time involved of using the actual vehicles and obstacles. Many companies are working to develop and combine vast and diverse databases of information to provide realistic driving simulations. These include roadway maps, road conditions based on weather models, friction coefficients based on tires and roadway construction information, databases, driving scenarios, and countless other sources of information. However, this information is only useful once the vehicle design has been completed.

To ensure the validation of automated vehicles while reducing the development cost and time, the industry needs a simulation platform that can test and retest the operation of the vehicle during the design of the vehicle’s systems. This would allow the test results to be fed back into the design, creating a closed-loop system that would improve the design, operation, and physical characteristics of the system in near real-time, as well as train and test the vehicle. Vehicle and systems designers would be able to experiment with different sensors, actuators, ECUs, and even processing solutions to test for optimal configurations and performance. For example, the simulations could be used to determine the placement of sensors, the best actuators and ECUs for response and power consumption, the selection of tires based on the vehicle weight and dimensions, or the battery displacement for the desired range of an electric vehicle (EV).

This simulation environment could also be used in the development of the system, from the individual components to the entire vehicle design. While models of many components, such as sensors, actuators, and ECUs, are available today, one model is often missing – the processing solution or System and a Chip (SoC). These SoCs are often designed for a broad range of applications and system requirements but there is a limit to the scaling of SoCs. As a result, there is a need for custom SoCs and the ability to design, verify, and validate them in the same simulation environment as the rest of the vehicle components and systems.

Stay tuned for more from this whitepaper. In the meantime, you can listen to this episode of the Autonocast in which Jim and I get into detail on the silicon requirements for AV ICs and SoCs.

Leave a Reply