“Wayve,” a relatively young British enterprise that focuses on Artificial Intelligence has recently announced that they’ve developed a more efficient way of training and controlling autonomous vehicles or AVs. It’s way more effective than the current methodologies that are using complex three-dimensional maps and expensive equipment.
To demonstrate the whole concept, they have released a video that shows a remodeled two-seater car (an electric Renault Twizy) as it navigates through the road smoothly, it seems that it has mastered things on its own.
The whole system works just like training a dog. Yes! It uses the reward and punishment technique which is also known as the positive and negative reinforcement approach. This means that the machine gets a certain form of reward when it behaves right and it gets penalties when it behaves otherwise.
The officials of the company explained that this type of approach works better and has already achieved great results so far. When the vehicle is treated like a child, it learns at a faster and more effective way when compared to the old methods used to train AVs to cruise through the streets by itself.
And in order to prove their argument, they’ve published a ten-page study showing the details of the experimentation they have performed.
The Issue on Maps
All the systems that exist which are developed to control AVs use complex 3D maps to aid in effective navigation. All around the world, different enterprises are competing to create the most precise map with the aid of highly-sensitive sensors and cameras that can take detailed photos of the streets and other stuff found on the road.
A good map is one that contains detailed descriptions and clear layouts of alleys, highways, motorways, alleys and every other place where humans and vehicles can traverse. And the vehicle itself needs to be equipped with sensors, cameras, and the needed gadgets in order for it to work smoothly with excellent accuracy.
But the main problem aside from the time it requires to develop such sophisticated systems is the frequency at which it needs to be updated. With all the changes that happen in the roads on a daily basis, these maps should also be updated especially in areas where construction is being done. That’s what needs to be done for AVs to work smoothly.
Moreover, most of the companies that develop these maps are only focusing on urban areas seeming to neglect the rural ones. The highly-congested conditions in urban streets make it harder for these AVs to navigate thus making any navigation system quite weak in these circumstances.
Cutting-Down on Technicalities
The new system developed by Wayve isn’t dependent on 3D maps and other technical stuff that current AVs operate on. For instance, most AVs today have multiple cameras on the front but the Renault Twizy used in the experiment only had one. Yes! just one camera while Tesla’s AV had eight!
That single camera is the one responsible for taking images and sends it instantly to the Graphics Processing Unit (GPU) located within the vehicle. The Graphics Processing Unit then runs the algorithm (rewards & punishment system) which is the one responsible for steering, slowing down or speeding up.
During the training, a human was seated in the driver’s area watching closely when the car is veering off the road. When it does, the human controls the steering wheel to prevent any undesired result and this becomes the form of penalty for the car. As the AVs time of driving on its own increased, it also got more of the reward (driving without human interference).
It took around 20 minutes for Twizy to learn the needed control parameters in order to cruise through a road that had a slight curve. Now that’s an achievement of miraculous extent. If this technology gets perfected, it’ll be a new way of controlling AVs that without the need of 3D maps. And when this happens, AVs will finally be available for everyone around the world to experience.