Reactive Collision Avoidance - Safe navigation in a dynamic cluttered environment
MetadataShow full item record
Collision-free navigation in unknown environments is an essential quality for any autonomous vehicle. In this thesis, a reactive collision avoidance algorithm is presented for vehicles constrained by a unicycle nonholonomic model in a multi-agent environment. The agents navigate independently in a decentralized manner, without explicit communication. Restricted forward speed makes the model suitable for vehicles with heavy linear constraints such as marine vessels and unmanned aircraft. The sensor model is given by an integrated representation of the environment where only limited sensing is required. A new braking rule is created to cope with typical multi-agent challenges such as oscillation and deadlocks. Through rigorous mathematical analysis, sufficient conditions for collision-free navigation is derived by reducing the number of agents. Tests, simulating thousands of cluttered environments, is presented including scenarios with both multiple agents and passive obstacles. The simulations prove that agents safely navigates the environment even when ignoring the strict conditions made in the mathematical analysis. Furthermore, the algorithm shows promising results when compared to other well known multi-agent reactive algorithms, such as the Reciprocal Velocity Obstacles. The main contribution of this thesis is a computational efficient reactive algorithm suited for a wide range of vehicles. By merging two existing algorithms and adding a new breaking rule, the result is a fast and safe multi-agent algorithm. In addition, a literature review is carried out to investigate alternative approaches to collision-free navigation and present the most relevant prior research in the field.