Developing a fully self-driving car is likened to landing on the moon. This is the technical, legal and even ethical challenge associated with putting artificial intelligence systems behind the wheel. Of all these problems, the need to be able to know where the car is at all times and be aware of its environment is one of the most important. And just snow can render the most advanced self-driving systems useless. For this reason, a team of researchers at the Computer Science and Artificial Intelligence Laboratory (CSAIL) at the Massachusetts Institute of Technology (MIT) has been working to develop a system that allows vehicles to map subsoil. Their approach uses Ground Penetrating Radar (GPR), which provides advanced detection capabilities. In this case it is another localized GPR (LGPR) developed by his MIT lab.
The usual solution to environmental awareness so far has been the use of video cameras and LiDAR systems. The latter is efficient when it comes to creating a 3D mapping of the environment, but laser technology cannot penetrate a blanket of snow, for example. Instead, GPR systems transmit electromagnetic pulses that reach depths of up to 3 meters and can detect the composition of asphalt and subsoil, the presence of roots and other elements. CSAIL leveraged these capabilities to integrate sensors into stand-alone vehicles and perform tests in closed snow-covered circuits.
this technology project We are still in the testing phase and have to overcome some obstacles. For example, the LGPR system used in our tests is 1.5 meters wide and must be placed outside the vehicle to function properly. However, researchers believe their approach could significantly improve the current capabilities of self-driving cars in the medium term.
A virtual driving academy for self-driving cars
Another MIT initiative in the area of self-driving cars is the development of a photorealistic simulation engine with limitless possibilities for learning reactions in virtual environments. A problem with the simulators used so far is that the data obtained from real human trajectories do not cover all possibilities. For example, it reacts less frequently to an imminent collision or lane intrusion by an oncoming vehicle. Now, the MIT researcher uses his simulator, called VISTA, to synthesize the myriad of trajectories a vehicle could follow in the real world.
The bottom line is collecting video data of human driving. Each frame is transformed into a 3D point cloud into which the virtual vehicle is deployed. Whenever the trajectory changes, the engine can simulate a perspective change and render another photorealistic scene with the neural engine. Each time the virtual car crashes, the system returns it to its starting point, which is considered a penalty. Over time, vehicles travel longer distances without colliding. Researchers then successfully ported this learning to real self-driving cars.
sauce: When
