Figure 1 – Types of sensors used for perception during autonomous driving
Thorough testing of the AV is very critical to avoid any misdetection / miscalculations that can result in life threatening consequences. On-road physical testing is time consuming, costly and requires lot of efforts. After driving millions of miles physically, a vehicle may not encounter some of the scenarios and there are less chances to repeat scenarios. In addition, several factors such as temperature, rain, fog, sunlight, etc. are not in the user’s control. In this case, ‘simulation based virtual test’ is the most feasible and essential during the development of the autonomous vehicle. Once the AV stack achieves a certain level of performance maturity in the simulator, it can be tested later with the vehicle on-road for a limited set of cases to reduce overall time to market. Therefore, Wipro is developing ‘SDV in a box’ - a global scale autonomous vehicle driving simulator used for testing and validating the perception and navigation algorithms of autonomous vehicles. It acts as a testing ground for autonomous vehicles before rolling it out on the road by simulating real time traffic scenarios.
AV simulator testing can only be effective if it is able to mimic the sensor behavior similar to the real world. The sensor models used to test an AV must accurately produce the data/signals generated by real sensors in real situations. Therefore, high fidelity virtual sensor models should be included along with realistic environment to capture the complexity accurately during testing and validation of perception, planning and control algorithms. Based on the level of fidelity requirement, virtual sensors can be modeled using one of the models shown in Table 1.