AgentDrive is a consolidated simulation framework for realistic vehicles simulation enabling testing of agent-based algorithms for route planning, navigation, cooperative driving, traffic optimization, and others vehicle coordination and cooperation methods.
AgentDrive covers several activities in the vehicles simulation and coordination domain. It goes from realistic vehicles modeling and simulation through multilevel control and coordination to planning and optimization.
Realistic simulation of the ground vehicles is based on the integration of agent-based simulation with the game physics simulation engine. The ground assets are simulated by a ray-casted model of a wheeled ground vehicle. In principle, the model can be of any shape and can use any number of wheels. The model is defined using several parameters including wheel friction, suspension stiffness, damping, compression, engine power, break force, and others. The fidelity of the physical model helps to anticipate real-world effects for future design of the control, planning, and cooperation mechanisms. E.g. the planning algorithm has to take in account details of the environment as friction, and momentum. The real position of an asset reflects the motion dynamic and introduces the uncertainty of the plan execution which the planner has to be aware of.
Planning, controlling and simulation of the vehicles is based on three layers architecture composed of (i) deliberative layer for high level coordination and planning, (ii) reactive layer for short horizon path planning, plan repairing and environment feedback processing, and (iii) simulation layer for physics simulation of the vehicle.
Multi-agent traffic simulation provides a realistic modeling with high fidelity continuous vehicle movement incorporating physical dynamics of the vehicles. The deliberative agent part of the system enables to implement and evaluate complex cooperation and coordination algorithms.
There are many different vehicles on a highway. Each of them acts somehow according to its logic. The logic determines behavior of the vehicle and reaction to critical situations. We distinguish cooperative and non-cooperative logic. Their main difference is in communication and interaction. Non-cooperative vehicles do not interact with other vehicles. They do not use negotiations with other vehicles. Cooperative vehicles are able to interact with the other cooperative vehicles.
Cooperative vehicles utilize the peer-to-peer collision avoidance algorithm originally developed for airplanes. Extended collision avoidance algorithm proved to be applicable in the highway traffic simulation in various scenarios. The vehicles than adapt their behavior with respect to agreed actions, e.g. in case of lane change the vehicles can adapt the speed in advance to avoid the need to break. The method helps the vehicles synchronize the speeds and improve the lane changing.
The vehicle routing problem and asset coordination is supported by the multi-agent solver based on task allocation. For the vehicle routing problem benchmarks the solver provides a solution with the quality of 81% compared to the optimal solution. It’s high applicability and low computational complexity allow us to deploy the solver not only in standard logistics domains (such as vehicle routing, pickup and delivery, time windowed problems, etc.) but also in wide scale of the dynamic routing and logistics scenarios such as cooperative navigation, dynamic frontiers exploration, undervalued cooperative tracking, cooperative area surveillance, etc.
The multi-agent task allocation solver supports routing and logistics problem solution in the tactical mission oriented projects for both ground and aerial asset scenarios. One of the strongest features is its high fidelity and scalability in dynamic scenarios with mixed types of assets.
Jiri Vokrinek (project coordinator, contact person), Antonin Komenda, Pavel Janovsky, Martin Schaefer, Karel Jalovec, Ales Franek, Jan Jiranek, Jan Harvalík