How Big Data fuels Tesla’s self-driven cars
An insight into the world of Tesla’s self-driven cars
Tesla, Inc. is an American automobile company founded in 2003 in Silicon Valley to replace the traditional fuel-powered cars with electric cars. Since then, Tesla, led by Elon Musk has become one of the most emerging brands in the automobile industry, especially when talked about electric cars. Today the company focuses on the two key areas electric propulsion and autonomous driving. Alongside automobiles, Tesla has also started building infinitely scalable clean energy generation and storage products like advanced battery and solar panel technology. The company believes in moving towards a zero-emission future and with less reliance on fossil fuels.
What keeps Tesla a step ahead of other electric automobile companies, is the way they make use of big data and artificial intelligence. It is a data-driven company. The business model followed by the company was to first create a large client base of people who like to try new technologies and then collect all the data generated by the cars and consumers. This collected data is analyzed with different algorithms and software. This data collection has helped them to improve and grow their system in every aspect. This is the reason the company term’s itself as a tech company and not an automobile company.
How the Big Data is collected?
Tesla provides a “tech package” to its car owners which include cameras and sensors that warn the drivers of accidents. Along with it, the sensors and cameras also collect the data like nearby buildings, pedestrian paths, road signs, and whatever that comes in its sight. Every Tesla vehicle running on the road is considered as a data point. Apart from collecting the road data their sensors even pick up the information of driver’s hand placement on the instruments and how they are operating them. Moreover, the driver’s reaction and action on unexpected events are also recorded which is used as an insight. With millions of cars running on the road, the company claims that it collects data of 15 million miles a day. This has helped Tesla to gather an enormous amount of data to feed to its machine learning and artificial intelligence models.
Applications of Big Data:
1. Training Auto-Pilot Model
The massive library of data collected from the cars running on the road is used to teach the neural network. A highly dense map is built using the data that includes everything from the average increase in traffic speed over a specific stretch to the accident-prone areas where action needs to be taken. The entire system is fed with this data using machine learning. The three major areas where data impacts the deep learning model of Tesla are discussed below:
· Object Detection
Tesla uses an object detection mechanism in which if the camera encounters an unrecognized object it takes a snapshot and uploads it to the cloud through Wi-Fi. Vehicles driving billions of miles makes it easy to source examples of rare objects. On encountering a rare object, the model is every time re-trained. The images of common objects are manually labeled.
In June 2020, Tesla has filed a patent titled ‘Enhanced Object Detection for Autonomous Vehicles Based on Field View’. This patent aims to improve the accuracy and efficiency of object detection through the images captured by cameras installed in vehicles in the future.
The basic concept behind the new mechanism is to focus on the objects in the image that have high computational requirements, while down sample less critical objects in the image. A Tesla vehicle uses a series of eight cameras to identify and recognize real-world objects. For example, a front-facing camera captures the image of the real-world location towards which the vehicle is heading. A part of these images may depict objects like pedestrians, vehicles, obstacles, etc. which are important in applications such as autonomous vehicle navigation. After determining the object, the particular field of view might be cropped from an input image and the remaining image may be down-sampled. This high-resolution and low-resolution portion of the input image may then be analyzed through an object detector.
The working of the model is explained using a flowchart shown in the below figure.
• Prediction
Prediction is the ability to forecast the actions and movements of objects like cars, cyclists, and pedestrians a few seconds ahead of time. Tesla uses its fleet of approximately a million vehicles as a resource to train the prediction model. Whenever the model makes a wrong prediction about the object, a snapshot is saved and uploaded to the training set. The model is then re-trained again using this data.
Tesla trains the prediction model in the “Shadow mode testing”. In shadow mode testing, the actions and reactions of a human, driving a car are compared with the predictions of the autopilot model. A car is being driven by the driver with autopilot mode on which receives data from the sensors but not taking the control of the car. A newly revised autopilot is also present at the same time. The new software makes decisions on how to drive based on the sensors. These decisions are compared with the decisions of a human driver or the older software. If the prediction does not match with the older software or with the human decision, the decision is marked, and the model is re-trained. For example, if the new software decides to zig whereas the older version decides to zag, or the new software accelerates while the human drives hit on the brakes, an attempt is made to compute how different the decision is and how important the difference is. In case of major differences, the portion of the incident is given to human beings to examine if the new software is making a mistake.
Tesla has the biggest business advantage in training its model over other companies as it has a fleet of approximately a million cars running on the road where the customer pays for that fleet attached with the sensors and trains their model.
• Path Planning
Path planning refers to the actions that a car takes while driving on autopilot mode which includes staying in its lane following the speed limit, overtaking a car, making turns on signals, nudging around a parked car, stopping for a jaywalker, etc. It is impossible to specify a set of guidelines using which the model decides any such situation. Tesla uses imitation learning to train its model for such circumstances. Imitation learning is training neural networks to emulate human behavior. In imitation learning, the model learns to predict what a human driver does by drawing correlations between what it sees via computer vision and what action the human driver takes. Tesla currently uses imitation learning to train for scenarios like how to handle steep curves or how to make a left turn at an intersection. In near future, it plans to use imitation learning for more tasks like how and when to change lanes on the highway.
2. Remotely providing updates and fixing issues
With the help of the cameras and sensors installed, the data-driven car can be tracked from the data center. This helps them to anticipate and rectify the issues remotely before they occur. Moreover, this also makes them capable of providing over the air (OTA) update. In 2014, the company provided overheating solution which reduced power fluctuation automatically without recalling the cars in the workshop. In 2015, the company provided an OTA auto-pilot update to 60,000 car owners which enabled their cars to drive by themselves.
3. Providing high customer satisfaction
The big data of the driver’s experience is used to analyze different performance parameters. These insights are then used by manufacturers to design vehicles that are safe and meet customer satisfaction. Not only from the sensors installed in the cars but also the company gathers data from the hosted online forums for complaints and feedbacks of the consumers. This data is further used for improvements and modifications in the current and new cars. Through this, the company doesn’t let their customers down and provides them with a pleasant and enjoyable experience.
Future Benefits
Tesla’s goal is to become the first company to reach full self-driving (FSD) automation. Currently, approximately a million Tesla cars are running on the road. This fleet of cars not only provides a billion miles of data but also simultaneously trains the existing model to reach full autonomy. Thus, the enormous amount of data keeps Tesla ahead of its rivals such as Waymo, Uber, and Lyft.
This data itself holds tremendous value. As per the reports of McKinsey and co., the market for the vehicle gathered data would estimate to be around $750 billion a year by the end of 2030. The company can also sell the data to other companies for navigation purposes as their maps are 100 times more accurate than any GPS. Moreover, Tesla is expanding its business and has now started producing solar panels and clean energy-based advanced batteries. The company makes use of its existing customer base data for sales and analytics of the new firms.
Challenges
The sensors and updates make the Tesla car safer. But at the same time, the sensors connected with the Internet makes it vulnerable to be hacked. To be aware of the potential vulnerabilities of the model, the company has a “bug bounty program”. In this program, the company rewards up to $15,000 to hackers and researchers who can hack or report any bug in their security system. The company has a team of top-notch security professionals who work on these vulnerabilities.
Keen Security Lab, a Chinese security firm hacked a Tesla car and was able to take control of brakes, side mirrors, windshield wipers and trunk while the car was driving at a distance from the hackers. They were also able to control the door locks, sunroof and vehicle’s lights while the car was stopped.
However, in just ten days Tesla provided a security update for the cars that would be affected. The update provided a cryptographic key to prevent future attacks.
Conclusion
Tesla’s business model makes it more of a data company than an automobile company. With the large fleet of Tesla cars running on the road, the company generates more than a billion miles of data in a year. To gather such a huge amount of data, other competitors still need several years. With the use of this data the company is successful in developing a semi-autonomous driving software and aims to develop a full self-driving (FDS) model in the future. The three major areas involved in the training of neural networks are object detection, prediction, and path planning. Moreover, the company also makes use of big data to remotely provide updates and fix issues, provide high customer satisfaction, and for sales analytics.
References
- About Tesla. (n.d.). Retrieved from https://www.tesla.com/about
2. Srikanth. (2019, August 24). How Tesla is Using Artificial Intelligence and Big Data. Retrieved from https://www.techiexpert.com/how-tesla-is-using-artificial-intelligence-and-big-data/
3. George Paolini. (2019, August 28). Tesla, the data company. Retrieved from https://www.cio.com/article/3433931/tesla-the-data-company.html
4. Albert Ahdoot. (2016, October 19). HOW BIG DATA DRIVES TESLA. Retrieved from https://www.colocationamerica.com/blog/how-big-data-drives-tesla
5. Staron, M., & Scandariato, R. (2016). Data veracity in intelligent transportation systems: the slippery road warning scenario. 2016 IEEE Intelligent Vehicles Symposium (IV)
6. Grigorescu, S., Trasnea, B., Cocias, T., & Macesanu, G. (2020). A survey of deep learning techniques for autonomous driving. Journal of Field Robotics, 37(3)
7. Shen, A., Phadte, R., & Joshi, G. (2020). Enhanced Object Detection for Autonomous Vehicles Based on Field View (US20200175326). The United States of America.
8. Klender, J. (2020). Tesla patent reveals Autopilot’s efficient method to enhance object identification. Retrieved from https://www.teslarati.com/tesla-patent-autopilot-enhance-object-identification/
9. Eady, Y. (2019, December 26). Tesla’s Deep Learning at Scale: Using Billions of Miles to Train Neural Networks. Medium. Retrieved from https://towardsdatascience.com/teslas-deep-learning-at-scale-7eed85b235d3
