Behaviour Mining for Collision Avoidance in Multi-robot Systems

Humans are very good at abstract spatial reasoning, both in the physical world and in virtual settings. For example, gamers controlling mobile agents in virtual environments can quickly identify potential collisions and effect evasive action. In a multi-robot system, avoiding collisions is a necessary and fundamental behaviour. Recent research has recognised challenges in scaling two-robot collision avoidance methods to larger populations. The work presented here takes advantage of humans' spatial reasoning skills, by collecting data from a simple 2D game where multiple robots move around in an enclosed arena and a human player prevents the robots from colliding with each other. Behaviours for avoiding collisions are learned by mining data logged during games. These mined behaviours are then deployed on robots in a non-interactive environment and compared to a fixed-threshold technique. Results show that mined behaviours provide an effective and flexible alternative.