Project topics and materials
Get Complete Project Material File(s) Now! »
Societal and economic impacts
In a broader picture, autonomous driving is expected to have ripple effects across a wide part of society and of the economy [5]. Indeed, self-driving cars promise to grant the same freedom of movement to almost anyone, by allowing elderly or disabled people to use a vehicle by themselves. Going another step further, some authors have proposed that autonomous driving would trigger a shift in vehicle ownership patterns, notably through an increase of car-sharing [9] which could, in turn, help decrease the amount of driving and parked vehicles in urban centers.
Although autonomous driving is largely praised for its potential benefits, some possible downsides – most of which are discussed in [6] – need also be mentioned. For instance, whether self-driving vehicles will effectively decrease traffic congestion is still debated, as temporarily reduced congestion could also prompt more people to use their own car. Since they can work or rest during their commute, people could also choose to live further away from their workplace, potentially leading to higher urban sprawl. Due to this induced traffic, the impact on public transportation is more uncertain: although removing the need for a driver may help decrease costs, a reduction in the number of passengers may require higher subsidizing as the poorest populationwill likely remain excluded from this progress. Similarly, although the ability to operate a vehicle without a driver can be highly appealing to the freight and taxi industry, the loss of employment for professional drivers should also be kept in mind, as well as potential effects in businesses such as insurance companies.
A classical architecture: the perceive, plan, act paradigm
Automated vehicles are a particular formofwheeled robots; as such, usual approaches to automated driving systems rely on the well-established perceive-plan-act paradigm (see, e.g., [14]). In this framework, a first perception layer is in charge of combining sensor data and potential prior knowledge such as mapping information into a suitable representation of the state of the ego-vehicle and its environment. The motion planning layer then computes a feasible and efficient trajectory for the ego-vehicle; finally, a control layer activates the vehicle’s actuators to follow this planned trajectory.
An alternative approach: end-to-end learning
With the dramatic increase of available computational power and the undeniable success of deep learning approaches in applications ranging from computer vision [15] to playing Go [16], a new approach called end-to-end learning has emerged for driving automation. This technique leverages machine learning to design a policy directly computing a sequence of actions (e.g., steering angle and acceleration) from raw sensor inputs (e.g., camera images), without requiring human-designed algorithms. Currently, two major techniques are being investigated for end-to-end driving. In a classical supervised approach (see, e.g., [17]), the algorithm is taught to imitate the behavior of actual human drivers in given situations, thus requiring a significant amount of training data recorded on real vehicles. More recently, the success of (deep) reinforcement learning algorithms in outreaching human [18] or even AI [19] experts in a variety of tasks without explicitly requiring training data has sparked a vast amount of interest in the artificial intelligence community. In this setting, an agent (the autonomous vehicle) learns through trial-and-error to choose actionsmaximizing a reward function defined by a human operator. For obvious reasons, reinforcement learning techniques are generally applied in simulations only, and it is unclear how training results can be transferred to the real world. These methods have the common advantage of considerably simplifying the development of driving algorithms, and mimic the way humans learn to perform complex tasks. However, they suffer fromseveral downsides that need to be addressed by future research.
First, current learning algorithms behave as black boxes with limited interpretability of failure scenarios and no method currently exist to provide behavior guarantees. Second, guaranteeing that trained models will respond well to unseen situations is still an active research topic. Finally, deep learning approaches have been shown to be vulnerable to (adversarial) attacks, where a tiny modification of the inputs leads to arbitrary changes in the outputs [20], which can result in security threats. For these reasons, we argue that using intrinsic structural properties of the underlying problem to guide or constrain learning algorithms is important for safety-critical applications; the results provided in this thesis may therefore constitute a first step towards this goal.
Table of contents :
Introduction
1 A primer on automated vehicles
1.1 Context,motivations and challenges
1.2 Motion planning in automated vehicles
2 Motion planning
2.1 Motion planning problems
2.2 Motion planning algorithms
2.3 Motion planning, homotopy and decision
3 Contributions and thesis outline
3.1 Cooperative driving
3.2 Autonomous driving
3.3 Towards practical implementation
3.4 Publications
I Cooperative motion planning
4 Priority-based cooperative decision-making
4.1 Multi-robot coordination andmotion planning
4.2 Priority and decision-making
4.3 Mixed-integer programming
4.4 Modeling priorities
4.5 Dynamic constraints
4.6 Chapter conclusion
5 Optimal coordination of robots along fixed paths
5.1 Time-optimal coordination
5.2 Problem modeling
5.3 MILP formulation
5.4 Simulation results
5.5 Some results for trafficmanagement
5.6 Chapter conclusion
6 Supervised semi-autonomy
6.1 Supervised driving
6.2 Supervision problem
6.3 Infinite horizon formulation
6.4 An equivalent finite horizon problem
6.5 Simulation Results
Contents
6.6 Chapter conclusion
II Motion planning for autonomous driving
7 Decision-free, near-limits motion planning
7.1 Aggressive motion planning
7.2 A simple second-order integrator model
7.3 MPC formulation for trajectory planning
7.4 Simulation results
7.5 Chapter conclusion
8 Classes of trajectories in autonomous driving
8.1 Motion planning for autonomous driving
8.2 State representation
8.3 Free space-time
8.4 Maneuver variants and homotopy classes
8.5 Chapter conclusion
9 Navigation graph
9.1 Free space partitioning
9.2 A guiding example
9.3 Mathematical results
9.4 Navigation graph and homotopy classes
9.5 Chapter conclusion
10 Graph-based decision-making
10.1 Decision-aware motion planning
10.2 Motion planning on the navigation graph
10.3 Global optimumsearch
10.4 Heuristics
10.5 Chapter conclusion
III Beyond simulations: bridging the gaps
11 Trajectory prediction
11.1 Trajectory and behavior prediction
11.2 Physics-basedMonte-Carlo estimation
11.3 A neural network approach
11.4 Chapter conclusion
12 A simplified implementation: velocity planning in the real world
12.1 Autonomous roundabout entry
12.2 Decision-making for velocity planning
12.3 Trajectory generation
12.4 Experimental results
12.5 Chapter conclusion
IV Appendices
A Complements on Chapter 4 A1
A.1 Computation of minimumbounding hexagons
A.2 Sub-timestep collision avoidance
B Complements on Chapter 5 B1
B.1 Helper constraints
B.2 Influence of time step duration on optimality
C Complements on Chapter 6 C1
C.1 Detailed demonstrations
C.2 Discussion on implementation
D Complements on Chapter 7 D1
D.1 Semi-infinite obstacles and local optima
D.2 Simulation model
D.3 Deriving a simpler dynamic model
E Complements on Chapter 8 E1
E.1 Unicity of the Frenet coordinates
F Complements on Chapter 9 F1
F.1 Demonstration of Theorem 6
G Résumés en français G1
G.1 Introduction
G.2 Partie I
G.3 Partie II
G.4 Partie III
