Autonomous Vehicle Navigation : From Behavioral...
In addition to flexible and bottom-up construction, multi-controller architectures can be formally analyzed to achieve reliable navigation in complex environments. This book reveals innovative control architectures that can lead to fully autonomous vehicle navigation in these challenging situations.
Autonomous vehicle navigation : from behavioral...
These three areas, combined, complement academic and industry work in perception, prediction, and planning. This workshop provides a platform to highlight recent efforts that advance behavior-driven autonomous driving, simulation, and ADAS with applications in human driver behavior modeling, multi-agent systems, and autonomous driving. Through invited talks, panel discussions, and paper presentations, attendees will become familiar with the latest research and network for new collaborations. Our jective for this workshop is to amplify the impact of behavior-driven autonomous driving research in both academia and industry, resulting in safer, robust, and confident autonomous vehicles.
One of the main goals of our workshop is to bridge the gap between the Multi-agent and Multi-Robot systems, Cognitive Robotics, and the Autonomous Driving community. The organizers and speakers are from top academic institutes (Stanford, Berkeley, CMU, Tsinghua) and industry organizations (Tesla, Waymo, Wayve, TRI) and have a lot of real world experience as well as a strong publishing record in dealing with these aspects. The organizers and speakers have also collected some challenging datasets on autonomous driving which are a great test bed for addressing research problems in cooperation, teaming, driver behavior, and game-theoretic aspects. Finally, we are proud to support great diversity in our team including age, gender, ethnicity, and location.
The specific goals of the workshop will be to discuss ideas around following topics: Perception in unstructured environments
Mapping and localization
Recognizing novel objects
Multi-agent trajectory forecasting
Multi-agent behavior prediction
Driver behavior modeling
Modeling human factors
Modeling human interactions in autonomous driving
Coordination and competition among multiple autonomous agents
Teaming in autonomous driving
Learning for multi-agent navigation
Theory for multi-agent systems
Multi-agent decision making
Reinforcement learning in autonomous driving
ADAS, simulation, and software-driven approaches
The path planning is one of the key issues of autonomous vehicle. In this paper, a behavior-based method is used for path planning. This method can manage complexity environment and is easy to design and test behaviors. An autonomous electric vehicle is tested in a driving course including the behaviors of road following, turning, obstacle avoidance and emergency braking. The test results show that the path planning algorithm can run in a real-time, while the unknown obstacles can be avoided and the autonomous vehicle can be driven toward the destination in a smooth path and continuous motion.
If you are interested in learning about the research and development (R&D) of autonomous vehicles, you really should read this book. It is an absolute must-read for people working in this domain; it is up to date (in an area that evolves very quickly) and full of references, and presents both a very good overview and a detailed discussion about some specific methods adopted in the development of autonomous ground vehicles.
Chapter 4 describes the reactive/cognitive hybrid architecture (hybrid-RC), which integrates reactive behaviors with high-level path planning approaches. Optimal path generation methods based on PELC (PELC* and gPELC*) are described. These path generation methods compete with other well-known methods from the literature, for example, RRT or A* based algorithms, but in the present case the author uses PELC as the base concept. In chapter 5, all of the presented concepts are put together in order to implement a real-world autonomous vehicle navigation system. The proposed system allows for the creation of optimal waypoints, but it also can use a pre-generated reference trajectory (waypoints). The proposed approach also allowed for the testing of a safe and reliable multi-vehicle navigation system (leader following). Real-world tests are demonstrated using VipaLab vehicles in the PAVIN platform for urban environment tests. Finally, chapter 6 presents a very interesting description of cooperative control of multi-robot systems, implemented using the proposed framework, that is shown to be adequate and can be adapted to allow this kind of application. In conclusion, this is a very interesting book, describing a well-proposed architecture and framework for implementing different autonomous vehicle applications (single or multi-vehicle).
Using PNs provides several advantages when defining the traffic rules and building the executive layer in the autonomous car application. They include: Flexibility, in which the traffic rules and behaviors are defined as PNs using the RG graphical user interface (GUI) , and they can be modified without the need to change the code; Module abstraction, which involves separating the modules interface with the PN from the implementation; Reduced development time, in which behaviors are implemented as PNs; Easy maintenance, which means tracing and debugging problems is easier when the system state can be seen by looking at the evolution of a PN rather than monitoring a set of variables; and Analysis and test, in which PN properties also make them good candidates for qualitative (un-timed models) performance evaluation and quantitative (timed models) performance evaluation of the car behaviors. Significant research has been carried out in this area for industrial applications [8,9] and also some in mobile robot tasks .
One of the main challenges when implementing a decision-making system for autonomous driving is the complex interaction between the car and the environment. Even in the case of modeling all the possible states and situations, the uncertainty introduced by the perception system  will include uncertainty to assess the current state of the vehicle and scenario. Furthermore, the presence of nearby external agents, such as other vehicles  or pedestrians , makes it especially challenging because their behavior is, in general, unpredictable.
Decision-making for autonomous driving is challenging due to uncertainty in the knowledge about the state of the vehicle and particularly the driving situation. This uncertainty comes from different sources, such as that introduced by the perception system  and the observers. Especially challenging is to estimate the continuous state of nearby external agents, such as other vehicles  or pedestrians , since their behavior is, in general, unpredictable.
Improving the observers means providing accurate sensors and better detection and tracking methods. Much research has aimed to anticipate future intentions of other traffic agents . The solutions proposed range from deterministic models  to different probabilistic models , such as Kalman Filters (KF) , Dynamic Bayesian Networks (DBN), and Hierarchical Dynamic Bayesian Networks (HDBN)  or Gaussian Process (GP) regression . Most statistic observers provide observation with a probability value or a probability distribution. Modeling these distribution probabilities is not an easy task and many of the models mentioned use some kind of reinforcement to learn some parameters of the estimated model or even the driving styles of other vehicles.
Another option to overcome the FSM state explosion problem is to use PNs. They have been used for different purposes in autonomous vehicles: evaluating the mission reliability ; designing and modeling a cruise control system ; modeling an intersection collision warning system , and modeling the cooperation of a human driver and an automated system . On the other hand, in the approach presented in this paper, PNs are used to model the desired behavior of the vehicle in different traffic situations.
If no pedestrian is detected, transition notPedestrian is fired, removing the token from watchForPedestrians and adding a token to the followPath place. While in this place, the local navigation module will keep following the lane provided by the map manager module. The system in this state is receptive to two events: pedestrian and close2Crosswalk.
When the transition close2Crosswalk is fired, a token is added to the reduceVel transition (Figure 6). The principal action associated with this transition is to command the local navigation module to reduce the speed. Still, the vehicle keeps following the path until the crosswalk is reached (crosswalkReached transition) unless a pedestrian is detected (pedestrian transition). If a pedestrian is detected, the pedestrian transition will be fired upon the reception of the corresponding message published by the event monitor module. A token will be added to the stop, and a message commanding local navigation module to stop in front of the crosswalk will be published. After the pedestrians cross and no pedestrians are detected, transition notPedestrian is fired, and the vehicle starts moving again.
Figure 19 shows the section of the UAH map (Figure 12a) where the crosswalk with pedestrian scenario takes place. This is a section of one of the road tests carried out in the UAH campus. The car starts from the right in autonomous mode with three people on board. For safety, if the driver uses the controls, the car switches to manual mode. As we pointed out before, the event monitor has to deal with several sources of uncertainty and can produce false events. However, it is important to notice that the behavior (PN) is only receptive to a few events in each instant. In order to save computing resources, the executive layer sends messages to enable and disable events according to the state of the PN. 041b061a72