Title: | Safe Motion Planning and Reinforcement Learning for Self-Driving Vehicles and Robots |
Subject: | Computer science, Software engineering, Robotics, Embedded systems |
Level: | Advanced |
Description: |
Overall Description Autonomous agents are systems that can move, function, and collaborate with others (e.g., humans or machines) without human intervention. Self-driving cars, drones, robots, and autonomous construction equipment are examples of agents in our real lives. For example, in Fig.1, a self-driving car is trying to merge onto the highway, which requires it to accelerate, turn, and act to possible emergencies, like parked broken down vehicles. Fig 1. An example of a self-driving car on highway, generated by CommonRoad Motion planning (MP) is about planning actions optimally so that the agent can achieve its goal efficiently and safely. In the example shown in Fig.1, MP is about planning the car's moving velocity and direction at different points so that it can merge onto the highway safely and timely. Machine learning is a popular way of training agents to behave in a certain manner s.t. reaching the goal area. Methods such as Q learning, deep learning, and Monte-Carlo tree search, are been used in solving the MP problems. Besides planning itself, safety is another important factor that one needs to consider in MP. Free of collision is one typical safety requirement. Never being stuck in a starving situation is another example of a safety requirement, which means that agents should never be waiting for each other s.t. a deadlock of waiting happens. To guarantee agents satisfy the safety requirements is extremely difficult but as important as planning itself because a minor error may cause catastrophic consequences. Formal methods, e.g., model checking, are techniques that are based on mathematics. They provide a way of proving that the target system is free of bugs in the sense that the system satisfies its requirements. So far, there is no such platform in academia nor in industries that provides an integrated way of simulation, training, learning, and verification of MP. In this thesis, you will be conducting research toward a platform for model-checked motion planning for autonomous agents. Content of the thesis
Contact Rong Gu (rong.gu@mdu.se) |
Start date: | |
End date: | |
Prerequisites: |
|
IDT supervisors: | Rong Gu |
Examiner: | Cristina Seceleanu |
Comments: | |
Company contact: |