Mitsubishi Electric advances vehicle control technology, also on general highways, for autonomous level 3 driving

For autonomous driving of ‘level 3’ or higher on general roads, Mitsubishi Electric has developed route generation and vehicle control technologies. It is a device that can enhance the accuracy of automatic driving on general roads and seamlessly follows the path produced with high precision in real time even when turning at a wide (deep) steering angle or when steep (yes) steering is necessary. It’s possible.

Achieved without AI (artificial intelligence) technologies being used. Aiming for acceptance after 2022 by domestic and international car manufactures.

In Level 3 autonomous driving, under such circumstances, the device works while tracking the situation inside the car. The driver takes over the operation if the machine was unable to execute the operation. Level 3 autonomous driving has now reached the realistic stage for single lanes on highways and motorways.

Honda will equip the completely improved “Legend” luxury sedan to be launched with a level 3 automatic driving capability in a single lane on expressways and motorways (Note 1) within FY2008. The type classification of the Ministry of Property, Housing, Transport and Tourism has already been obtained and states, “This is the first time in the world to specify the type of a level 3 self-driving car” (Ministry of Land, Infrastructure, Transport and Tourism). In the future, the entrance of domestic and international car suppliers is expected to accelerate .

Notice 1) The “ProPilot 2.0” of Nissan Motor and the “New Generation EyeSight” of SUBARU know automatic driving for several lanes on highways and motorways.
In terms of infrastructure, autonomous driving at level 3 is feasible, but it is restricted to “level 2” driving assistance in terms of service.

General road level 3 that includes specialized functions
To track the situation inside the car, the Level 3 autonomous driving system uses sensors such as cameras, millimeter-wave radar and LiDAR (laser scanner) and uses high-precision maps and GNSS (Global Positioning Satellite System) to monitor the environment of the vehicle. Place grip. Based on this information, it is important to construct an automated driving route and monitor acceleration / deceleration and steering so that the route can be correctly followed.

However in relation to highways and motorways, there are many items to be found on general roads, such as cars, pedestrians and bicycles (bicycle drivers). The road setting, such as intersections with low visibility and curves with narrow roads and wide turns, is more challenging than highways and motorways.

More advanced functions are needed not only for sensors, but also for route formation and vehicle control in order to realize Level 3 autonomous driving on general roads. For eg, even if a path is generated on the basis of sensor knowledge, GNSS, or a high-precision vehicle control chart, the real driving status can vary dramatically from the target route. Furthermore, because of the delay in the actuator response time, there is also a concern that the precision of route tracking usually decreases.

Achieves smooth and highly accurate automatic driving
 To solve these problems, Mitsubishi Electric has developed software technology for route creation and vehicle control. In the first route creation technology, sensors, GNSS, and high-precision maps are used to detect objects around the vehicle and to grasp the position of the vehicle on the map. Then, using a uniquely developed method called “particle filter (PF)”, the optimum route is generated in consideration of the environment around the vehicle.

 Specifically, the technology (PF) that creates vehicle positions over time and connects the positions closest to the target from among them to create route candidates, and expands the route candidates created by PF. However, we combined a technology called “RRT (Rapidly-exploring Random Tree)” that selects the optimum route from among them.

 The second vehicle control technology is accurate and stable by predicting the movement of the vehicle up to a few seconds ahead and controlling steering and acceleration / deceleration so that the movement is closest to the path created by the PF and RRT. To be able to follow. For example, even when traveling on an S-shaped curve having a large curvature, it is possible to prevent the vehicle from swelling or meandering due to a delay in steering or acceleration / deceleration.

 In some cases, domestic and foreign automobile manufacturers are working to apply AI technology not only to the recognition process but also to the route creation and vehicle control processes in order to realize Level 3 autonomous driving on ordinary roads. is there. In response, Mitsubishi Electric said, “We used a sampling-based planning method for route creation and a nonlinear model predictive control method for vehicle control. We do not use AI technology.” (Autonomous control at the company’s Advanced Technology Research Institute) Mr. Masaya Sakai, who is the automatic driving system group manager in the system development project).

Confirmed performance with a test vehicle equipped with new technology. Mitsubishi Electric conducted several driving experiments using a test vehicle equipped with the developed technology. The test vehicle was equipped with an ECU (electronic control unit) for autonomous driving, and this technology was incorporated into the ECU.

 One of the driving experiments is to avoid a vehicle (obstacle, the same applies hereinafter) stopped in the driving lane and the overtaking lane on a straight road with two lanes on each side. While traveling at a vehicle speed of 80km / h, the sensor detects obstacles in front, and while decelerating, the steering is automatically operated to avoid two obstacles. It was confirmed that obstacles could be avoided while accurately following the created route.

 An experiment was also conducted in which the vehicle traveled at a vehicle speed of 30 km / h on a frozen road (ice road) and avoided obstacles in front of it. In a vehicle not equipped with the new technology, if the steering is operated while decelerating in an attempt to avoid an obstacle, the rear part (tail) of the vehicle may be greatly swung outward.

 On the other hand, in vehicles equipped with new technology, the movement of the vehicle is controlled so that the grip of the rear wheels is not lost (so that the rear part of the vehicle does not swing outward), and obstacles are followed while following the created route. Was avoided. Mitsubishi Electric plans to improve its technology to handle more complicated driving scenes.

Predict 5 seconds ahead and create a route
 Among the technologies installed in the test vehicle, the proprietary technology called PF Note 2) mentioned above was used to create the route . Using this technology, the route is automatically created so that the vehicle approaches the target state (such as being located in the center of the lane).

Note 2) PF is a state estimator that approximates the conditional probability distribution by a group of data called particles. It is used to estimate the unmeasurable state based on the measured value obtained from the sensor.
 First, a plurality of particles (particles) representing the state (position, speed, orientation, etc.) of the car are generated, and the number of particles is increased or decreased according to the likelihood (likelihood). Likelihood is a statistical term, the degree to which an estimate can be judged to be valid. Decrease particles with low likelihood and increase particles with high likelihood.

 Next, the previous process is repeated several times, starting from the increased or decreased particles. Finally, the weighted average positions of these particles are connected to create path candidates

Vehicle control method that can reduce the amount of calculation
 Based on the route created using PF and RRT, vehicle control uses a method called nonlinear model prediction control method to integrate and control the vertical and horizontal movements of the vehicle. Specifically, the state (position, speed, etc.) of the vehicle with respect to the control input value (degree of acceleration / deceleration, steering angle, etc.) is predicted. Adjust the control input values ​​so that the predicted vehicle condition matches the target route

 By repeating this process, the control input value is optimized so that it best matches the route created by RT and RRT. As a result, stable tracking of the created route becomes possible. Since the amount of calculation is small in this control method, the calculation time can be shortened. “We predict and control movements 2.5 seconds ahead,” Sakai said.