Artificial intelligence developments are set to fundamentally transform mobility, whether it is mobility of weapon payloads, supply deliveries, urban commuters, warehouse goods, delivery packages, or intercontinental bulk shipments.
The chief problem in making aerial systems, ships, vehicles and robots autonomous by replacing human pilots, operators, and drivers with artificial intelligence, is that of machine perception. An autonomous system's computer needs the ability to precisely recognize other vehicles, humans, signs and markings, trees, buildings, terrain, etc. And that, too, in poor environmental conditions, such as in the darkness of night, in rain and in snow.
The problem is nearly impossible to solve with traditional rule-based control algorithms. Instead, neural-networks and machine-learning methods need to be used. In these methods the computer is trained rather than programmed.
Since these machines must make sophisticated decisions in highly complex environments, the number of use-cases and scenarios that they must be designed and tested for is prohibitively enormous. For instance, analysts estimate that autonomous road vehicles will need to be driven through billions of miles of road-tests in order to make them as safe and reliable as human driven vehicles. In contrast, with six years on the road, Google's self-driving car project recently completed 2 million miles of road tests. At this rate, it will take millennia to complete the road tests needed for the development of autonomous vehicles.
This impossible task can only be accomplished with the speed and cost economy of engineering simulation. With simulation, thousands of tests can be virtually conducted within the time and resources needed for a single physical test, thus greatly accelerating the development of autonomous systems.
Furthermore, safety and reliability of autonomous systems are of major concern. An inaccurate sensor, a software bug, or a malfunctioning actuator on these intelligent machines can easily cause expensive equipment failure, unintended human injuries or fatalities, and mission failures. Therefore, electronics, software and hardware for these machines must be thoroughly tested during the engineering design process with simulation for ensuring safety and reliability.
Simulation accelerates autonomous systems development in six areas:
- Mission Scenario System Simulation;
- Software and Algorithm Modeling and Development;
- Functional Safety Analysis;
- Sensor Performance Simulation;
- Electronics Hardware Simulation;
- Semiconductor Simulation.
Mission Scenario System Simulation
Comprehensive mission scenario simulations can be conducted with a system level behavioral model of an autonomous system. Such a model includes all sensors, control systems, drive systems and vehicle body, placed in situ in a virtual scenario environment comprised of terrain, roads, building, humans, road-signs, etc. In this simulated environment, thousands of mission scenarios can be evaluated rapidly, to test whether the vehicle's sensors, control algorithms, and drive systems perform as expected under various situations.
Simulating the complete mission scenario requires comprehensive system simulation. The first step is a world model comprised of virtual terrain, roads, buildings, humans, other vehicles, etc. The vehicle being studied – referred to as the ego-vehicle – moves in this virtual world.
Next, the sensors on the ego-vehicle are modeled. Radars, lidars, ultrasonic sensors, cameras and other sensors on the vehicle observe the virtual world surrounding the ego vehicle and generate simulated sensor signals. The sensor signals are then passed to signal processing, sensor fusion and control algorithm models. The algorithm models make decisions about altering the vehicles speed and orientation. The control decisions are then passed on to virtual models of the actuators such as thrusters, power-train, and robotic limbs that control the vehicle's movement.
The vehicle's movement is computed by the vehicle dynamics module of the system simulation which precisely predicts the movement of the vehicle, taking it to a new position in the world model. The entire control loop then re-repeats endlessly step-by-step until the mission scenario is completed.
Various parameters can be tested in such mission scenario system simulation with speed and ease. For instance, one can conduct a what-if study to see the effect of an unexpected failure of one of the vehicle's sensors. Additionally, such scenario simulations are highly valuable for regression testing of software and algorithms. The speed, cost economy, accuracy, and automation of scenario simulation makes it an indispensable tool for repetition of a pre-defined set of regression tests.
Software and Algorithm Modeling
Just as in hardware development, simulation has a key role to play in software development. Developing and testing signal processing routines, sensor fusion algorithms, object recognition functions, control algorithms, and human-machine inter-face (HMI) software, with model-based software development techniques makes the software robust, less error-prone, and safe.
Model based software development and verification tools such as the Safety Critical Application Development Environment (SCADE) more than double the speed and productivity of developing software applications complying with safety standards such as the DO-178C A-level in aviation and ISO26262 ASIL-D level in automotive. For instance, Airbus noted that their software development time reduced by a factor of 3 by using SCADE as compared to a process without automatic code generation. More importantly, Airbus never experienced any bug in their flight control software that was produced with automatic code generation.
Functional Safety Analysis
The complexity of autonomous systems greatly multiplies possible failure modes failure cascade paths. Since autonomous systems inherently have safety implications, any failure can easily be catastrophic, even fatal. Conducting functional safety analyses of such complex systems, is tedious, error prone, and vulnerable to gaps and flaws. Automated functional safety analysis tools are, therefore, essential to ensure safety of autonomous systems.
Having a functional safety analysis tool in a simulation platform streamlines and expedites virtual testing of possible failure modes to evaluate their criticality and to develop counter-measures. Use cases include the utilization of simulation to provide evidence for earlier assumptions in functional and technical safety concepts as well as the ability of understanding potential failure modes through simulation.
Sensor Performance Simulation
Sensors are key new components that need to be developed for autonomous systems. Simulation uses high-fidelity physics to predict the performance of sensors such as radars, communication antennas, and ultrasonic sensors. For instance, simulation predicts radar patterns and gain in specific mission scenarios, eliminating expensive and time-consuming physical testing. Further, simulation computes the changes in performance of a radar when it is mounted on a vehicle, and when it operates in rain or snow, providing precise insights into real-life radar operation, at a fraction of the cost and time needed for field tests.
3D electromagnetic field solvers based on the Finite Element Method (FEM) and Shooting-Bouncing Rays solvers (SBR) are used for performing high-fidelity simulations of automotive radars. These simulations accelerate four radar development aspects, as follows:
(a) Isolated Radar Simulation: Here the radar antenna(s) and radome are simulated as placed in free space. Rapid parametric studies are conducted in such simulations to optimize geometric and material design of antennas and radomes.
(b) As-Installed Radar Simulation: Here the radar is simulated as installed on a vehicle to determine the degradation of radar performance due to obstructions caused by the vehicle's neighboring components.
(c) In-Environment Radar Simulation: Here the performance of a radar is simulated in a large, realistic environment comprised of other vehicles, buildings, humans, trees, etc. Given an input signal at the radar's transmit antenna(s), the high-fidelity physics simulation computes the output signal at the radar's receive antenna(s) based on what the radar “observes” in the virtual environment where it is placed.
(d) In-Mission-Scenario Radar Simulation: Here, reduced order models (ROMs) of high-fidelity radar simulations are used to create fast-executing, yet high accuracy, models of radars that can be used in mission scenario simulations described above.
Electronics Hardware and Semiconductor Simulation
Autonomous vehicles contain a host of new electronics hardware in the form of radars, lidars, cameras, other sensors, communication systems, signal processing systems sensor fusion boards, artificial intelligence computers, controllers, actuators, and HMIs (human-machine interfaces). The components need to be designed to withstand electrical, thermal, vibrational, and mechanical loads without failure over the lifetime of the vehicle. Simulation greatly expedites the speed of testing designs and provides physical insights that enable engineers to optimize electronics components and make them robust.
High-fidelity 3D physics-based simulation helps to analyze various physical phenomena across electronic packages, boards, enclosures, and systems, such as power optimization, power integrity, electrostatic discharge (ESD), electromagnetic interference/ electromagnetic compatibility (EMI/EMC), thermal, and structural reliability.
Simulation and modeling tools also help chip designers with accuracy and performance needed to reduce power noise and improve reliability of ICs that are being specially developed for autonomy applications. The challenges designers face are higher temperatures from higher current densities, self-heat, electromigration (EM) and electro-static discharge (ESD). An increase in temperature of 25°C typically leads to 3× to 5× degradation of the expected lifetime of devices. High-fidelity physics-based simulation provides accurate thermal analysis with robustness and connectivity checks, power and signal EM checks, full-chip ESD analysis, and effects of self-heating.
The more intelligent we make our machines and the more autonomously we let them operate in the open world, the more the operating scenarios that we must consider in designing and testing these machines. Such scenarios can easily number in millions, if not billions. Physical testing is prohibitively resource intensive and, therefore, impossible. Yet we can develop viable unmanned systems, autonomous vehicles, and robots with the help of engineering simulation that can virtually test a vehicle in thousands of operating scenarios in a fraction of the time and cost needed for physical testing.
This article was written by Sandeep Sovani, Ph.D., Director, Global Automotive Industry, ANSYS Inc. (Canonsburg, PA). For more information, Click Here .