Companies such as Waymo, Toyota and Uber have decided to test autonomous vehicles on the streets of Phoenix, Mountain View and other cities. But, interestingly, they will not test on actual, real streets, rather they will test them in a computer simulated virtual reality. A few engineers think that it will be better to make mistakes and learn in virtual reality rather than on actual streets.
This way, flaws can be identified without actually imperilling the lives of people on the streets. If there is a fault on the simulated drive, engineers can alter and programme the software settings accordingly and lay down new behaviour setting.
Waymo, the self-driving car company that rose out of Google, is expected to demonstrate its simulator tests in California’s Central Valley. Not only this, researchers are also coming up with ingenious methods that would make it possible for cars to adopt new behaviour from these simulations alone. Such virtual training and modification would be quicker than whatever will be formulated by human engineers using external software code.
Google had developed its first autonomous cars by engineering the software bit by bit, minutely coding each and every behaviour. Now days, through machine learning and advancement in computing power, autonomous automakers are creating intricate algorithms, which can process information on their own.
However, such advanced technology has been treated with scepticism. It can be challenging to understand the behaviour exhibited by these algorithms and their cause. As compared to humans, they analyse more information. Nevertheless, machine learning will prove to be the lifeline for autonomous vehicles.
After a decade of research, testing and development, Waymo — under Google — is set to offer autonomous public rides in Phoenix, Arizona. These taxi services will not require a human to be behind the car’s wheel. However, any sort of autonomous vehicular transit would preferably happen in a small area, with less people and little to no rains.
Through continued testing and development, these cars should be able to handle more challenging scenarios eventually. By incorporating new sensors, complex algorithms and neural networks, these autonomous cars will be able to examine data and learn tasks in a much more versatile and efficient manner. Algorithms can train the cars to identify people, traffic signals, lane markers, etc.
However, researchers and developers will be required to gather and label data that describes each and every conceivable situation and scenario, which seems like an impossible task. That is precisely when simulations come to the rescue. Waymo recently revealed a roadway simulator called Carcraft. Using this simulator, Waymo can test cars in a much better manner than possible in the actual, real scenario.
Simulations are very similar to the real world, which enables the systems to be reliably trained. Apart from that, researchers have complete control over a simulator as they do not need to label images.
World’s leading artificial intelligence labs, such as OpenAI, DeepMind and the Berkeley AI Research Lab, have come up with a new method called reinforcement learning—complex algorithms enable machines to learn tasks in a simulated setting through in-depth trial and error. If machines can learn how to navigate in a virtual world, then they can do well in the real world too.
However, algorithms are always vulnerable to learning faulty behaviour patterns. In order to counter that, Waymo and Toyota are incorporating hand-coded software using conventional ways as well.