What are Self-driving Simulations?
The Encyclopaedia Britannica defines a computer simulation as “the use of a computer to represent the dynamic responses of one system by the behaviour of another system modeled after it.” Airline pilots train on simulators, while in the automotive industry driving simulations are used to optimize ride experience and improve engine performance. With the emergence of autonomous vehicle technologies, the development and deployment of effective self-driving simulations has become an industry priority.
Self-driving simulations collect data to improve autonomous vehicles’ algorithm training capability, sensor accuracy and road data quality. It’s been proposed that autonomous vehicles should log 18 billion kilometers (11 billion miles) of road test data to reach an acceptable safety threshold. This would however require a fleet of 100 vehicles running on roadways 24/7 at a constant speed of 40 kph (25 mph) for 5 years. Self-driving simulations are an ideal solution, enabling continuous and unlimited data collection in the digital environment with relatively low operational costs.
Self-driving simulations have advantages in milage data collection efficiency, road condition dataset diversity, and sensor corresponding data accuracy.
Why Self-driving Simulation?
Data collection and operation efficiency
Self-driving simulations can boost the speed of data collection to reach mileage accumulation targets while reducing fleet operation costs. Waymo and Baidu have made use of self-driving simulation to speed up self-driving development.
- Alphabet subsidiary Waymo has accumulated the most mileage of all self-driving companies, with a February 2018 total of nine years and five million miles. Waymo simulations created over 2.5 billion self-driving miles in 2016 alone, 500 times the fleet’s real road mileage. The Waymo simulator transforms real world scenarios into virtual formats and runs 25,000 virtual cars simultaneously. The massive data flow from this simulation process assists engineers in locating bugs and adjusting models efficiently.
- Baidu’s self-driving development platform Apollo also works to speed the accumulation of self-driving mileage, and can simulate a million driving miles daily. The platform allows partners to share selected self-driving data with each other to optimize solutions together. In addition, Apollo provides most of the infrastructure, data and functions required for self-driving development, so engineers don’t need to create everything from scratch.
Dataset diversity and environment complicity
Self-driving simulations can also add more uncertainty to a dataset to increase the responsiveness of the system. They can produce a variety of scenarios to test and improve vehicle performance under different conditions, for instance in severe weather, heavy traffic environments, and various distinct scenarios. Truevison.ai and AirSim are two leading solutions in this field.
- Truevison.ai is a company working on photorealistic self-driving simulation. It can provide simulations with rich details in a variety of environments such as highway, city, and mountain road scenarios. The simulator is also able to easily introduce weather conditions such as rainfall, snowfall and fog. These changeable factors can increase the diversity of dataset.
- AirSim has a very detailed 3D urban environment that contains a wide variety of dynamic scenarios such as traffic lights, parking, construction sites and much more. The simulation includes more than 12 kilometers of roads and 20 city blocks. This virtual city environment can help engineers increase the robustness of their self-driving systems.
Multiple raw data and sensor accuracy
Sensors are the eyes of the self-driving car. Multiple sensors can increase the accuracy of input data and protect the vehicle from misidentification or malfunction. Self-driving simulations provide multiple signals simultaneously, such as camera, LiDAR, radar and more. A given object in the virtual environment will thus be detected by different sensors, and these signals will validate each other to increase accuracy. RightHook and Cognata are two simulators providing multiple self-driving sensors.
- RightHook’s self-driving simulation can generate multiple signals such as LiDAR and radar to create a virtual environment on an HD map with no other input, thus reducing the cost of scene reconstruction. RightHook’s multiple sensors can expand the training dataset and improve system performance. Engineers can customize sensor data and even adjust the sensors’ location on the vehicle for specific purposes.
- Cognata focuses on deep learning self-driving simulations. The company collects real sensor data and builds its virtual environments and scenarios accurately with the help of deep learning. The various sensor data inputs can validate each other to provide more precise results.
Limitations and Expectations
The limitations of self-driving simulations should not be overlooked. Currently, there are two main weaknesses: lack of emergency situation scenarios, and potential consequences of differences between real and simulated data.
Emergency situations are still hard to simulate as each real world accident is unique. Although vehicles can learn general driving operations through simulation, it is impossible to predict every single emergency situation. For example the May 7th, 2016 fatal accident involving a Tesla on autopilot occured because the system failed to distinguish a white truck against a bright sky. Although this is not a rare scenario in the real world, the simulator had not covered it.
This differences between real and simulated data is another issue that could negatively affect system performance, as the full consequences of these differences remain unclear. Engineers are hard pressed to determine what type of data leads to an accident due to unexplainable features of the algorithms. For example, real world pedestrians with disparate clothing and posture profiles cannot all be reflected in the simulator, and this might reduce a self-driving vehicle’s ability to identify pedestrians.
In order to overcome these drawbacks, it is important to simulate more scenarios to make abnormal data traceable. SynCity is a Dutch company developing an advanced simulator that aims to present a wider range of scenarios, including other vehicle misbehaviour and various emergency situations, in order to optimize algorithms.
In the words of Toyota Research Institute Chief Executive Gill Pratt, “Simulation is a tremendous thing.” The correct understanding and prudent application of self-driving simulations are essential to safely accelerating R&D in the autonomous vehicle industry.
Reference:
https://t.qianzhan.com/caijing/detail/180205-888d2fbf.html
http://technews.cn/2017/11/30/waymos-fleet-reaches-4-million-self-driven-miles/
http://apollo.auto/platform/simulation.html
https://www.leiphone.com/news/201801/RQ57jcE145rzKH2o.html
https://www.nytimes.com/2017/10/29/business/virtual-reality-driverless-cars.html
https://www.leiphone.com/news/201711/x8bdcKBiK7eQCM9s.html
https://www.leiphone.com/news/201712/a6x9WnVVBM76FOWt.html
https://www.leiphone.com/news/201706/FZcktWIYi3NrmsO8.html
Analyst: Alex Chen| Editor: Michael Sarazen
0 comments on “Simulations Pave the Road for Self-Driving Technologies”