The MapLite system combines global positioning systems (using only the most basic topographic maps from OpenStreetMap), as well as lidar and IMU sensors for monitoring road conditions.
The system can obtain the geographical position information of the car through GPS and use this information to identify the 'local' target in the final destination and the car's field of view. Then, this device can use its on-board sensors to generate points for reaching these places. The path, and the use of laser radar to estimate the edge of the road. This system is prefabricated with a number of general-purpose, parameter-based models that allow the car to know what to do at an intersection or on a special road.
This system may help prevent future accidents, such as the recent deaths involving Uber's self-driving cars.
To date, many self-driving cars tested on actual roads often require well-marked 3D maps to identify speed limits, lanes, and signs. But on rural roads, these markers are often absent.
Therefore, the MIT CSAIL team developed a method that allows self-driving cars to recognize and predict the local environment, rather than relying on 3D map data.
Teddy Ort, a graduate student in the computer science and artificial intelligence lab, said: 'This 'mapless' approach has not really been achieved because it is often difficult to achieve the same accuracy as a detailed map. Sex and reliability. Systems like this can be navigated through on-board sensors, which shows the potential of self-driving cars to actually handle the few roads that technology companies have not painted.'
The researchers drafted a paper describing the system that will be released at the International Conference on Robotics and Automation (ICRA) to be held later this month in Brisbane, Australia.
To test the system, they were equipped with lidars, sensors and MapLite for the self-driving car Toyota Prius. The Prius car passed 'saw' more than 10 feet (about 3 meters) ahead of the road, multiple strips in Devon, Massachusetts. Unpaved rural roads are successfully navigated.
Ott explained that their system is different from other 'no map' methods, which use machine learning to train the system.
The researchers stated that their system could not handle the rapidly changing environment, so it was impossible to navigate on the mountain.
Ott said: 'I can imagine that future self-driving cars will always use 3D maps in urban areas, but when we are asked to start from secluded roads, these vehicles need to be like humans, and they have never seen them before. Driving on a strange road. We hope our work is a step in this direction. '