| 研究課題/領域番号 |
21F21362
|
|---|
| 研究機関 | 東京大学 |
|---|
研究代表者 |
中野 公彦 東京大学, 生産技術研究所, 教授 (90325241)
|
|---|
| 研究分担者 |
CHENG SHUO 東京大学, 生産技術研究所, 外国人特別研究員
|
|---|
| 研究期間 (年度) |
2021-11-18 – 2023-03-31
|
| キーワード | dynamics / control / automated vehicle / neural networks / decision making |
|---|
| 研究実績の概要 |
This project aims to investigate chassis dynamics domain control for automated vehicles (AVs). A decision-making module provides proper driving commands for chassis actuators to track, so the first step of studying chassis domain control is to design or employ a decision-making module. Here we have investigated a convolutional neural network-based approach to generate both longitudinal and lateral driving maneuver commands.
This project has established a simulation platform based on the simulation of urban mobility (SUMO) to collect driving data. Then, a spatio-temporal image representation approach was proposed to depict full-scale traffic information in motion-sensitive area around the ego car. After that, bulks of motion images together with driving maneuver data were collected and divided into training and test dataset based on the SUMO platform. We have designed a network architecture with the convolutional neural networks and long short-term memory layers to extract both spatial and sequential traffic representations and learn underlying driving operation features. The deep-learning-based decision-making method has been trained and tested by the collected datasets from SUMO, and the current results validate the effectiveness of the proposed scheme. This part of research work has been submitted to IFAC Mechatronics -MoViC2022 Conference.
|
| 現在までの達成度 (区分) |
現在までの達成度 (区分)
2: おおむね順調に進展している
理由
Currently, we are attempting to build a training dataset from naturalistic real-world traffic trajectories instead of the SUMO platform. We will choose some proper open-source trajectory datasets collected in the real world. We are searching for what datasets are available and suitable for this project. Furthermore, this project is still going on writing codes to process naturalistic datasets. The flowchart of processing data has been designed. Firstly, read each row iteratively, meanwhile build suitable data structures to contain sequential snapshots of all time stamps and trajectory data of all vehicles. Then, put the new data frame into the two data dictionaries and sort all elements by time stamps and vehicle IDs.
On the other hand, this project now is researching chassis global dynamics modeling of AVs. We begin with a comprehensive analysis of the vehicle cyber physics system, which covers nonlinear dynamics modeling and spatio-temporal signals. Vehicle longitudinal, lateral, and vertical dynamics motion is being analyzed as well as wheel dynamics. To investigate the correlation among different vehicle dynamics signals, this project currently records some simulation results of typical test scenes and seeks to find out their inherent statistical relations.
|
| 今後の研究の推進方策 |
For the decision-making research part, the next step is to construct the training dataset and test dataset from naturalistic traffic data in the real world and then train our proposed deep-learning-based decision-making method. The research fellow plans to write one paper about these research results and present them at one academic conference.
For chassis domain control, the safe and stable driving boundary is one fundamental basis. After current work on chassis global dynamics modeling, this project plans to explore the unclear boundary of safe driving, this project will propose the mathematical definition and rigorous design of the performance metric and investigate the specific range of metric values. A rigorous metric will be developed to quantify AVs’ inherent safety.
Considering multiple actuators of vehicle chassis, this project will propose a multi-agent cooperative method. Then, chassis steering, braking, and driving components will be coordinated by the designed cooperative strategy with the same goal of vehicle performance equilibrium optimal. To test the performance of the chassis domain control proposed by this project and validate its efficiency, this project will construct one hardware-in-the-loop test platform and design typical traffic scenarios, and then carry out some typical experiments. This project plans to publish the above research works in top journals.
|
