Cooperative Perception

There are many features in a soccer game that are important to improve game playing. Most of them can be determined by a single robot, but sometimes (e.g., due to occlusion) cooperative perception is required. In any case, "double-checking" a feature by more than one robot is supposed to improve the estimation of the feature attributes. Examples are: determining the ball position from the estimates of several team robots, as well as from different sensors within each robot (in our case, the front and up cameras); improving the estimation of a teammate posture, using its own estimate and its observation by the teammates.

We have developed sensor fusion algorithms to estimate ball location, other robots, goals and field boundaries from the observations made by several sensors, either onboard the same robot or distributed over the team robots. This lead to a distributed world model, where information is stored in the distributed blackboard. The sensor fusion method is based on Durrant-Whyte's work, and allows two sensors to disagree. In such a situation, the observations are not integrated and each robot uses its local estimate. The work is described in IAS-8 conference best poster paper "Bayesian Sensor Fusion for Cooperative Object Localization and World Modeling". Recent work has extended this older work by tracking the ball in 3D with particle filters and fusing the ball position and velocity information from different robots in the world frame, taking into account the robots self-localization uncertainty and possible disagreements.