Navigation

The ability of the robot to self-localize, i.e., to autonomously determine its position and orientation (posture) within its surrounding environment, as well as its capacity to move from its current posture to a desired new posture is one of the robot most important features. Once a robot knows its posture, it is capable of following a pre-planned virtual path or stabilizing its posture smoothly. If the robot is part of a cooperative multi-robot team, it can also exchange the posture information with its teammates so that appropriate relational and organizational behaviors are established.

In robotic soccer, these are crucial issues. If a robot knows its posture, it can move towards a desired posture (e.g., facing the goal with the ball in between). It can also know its teammate postures and prepare a pass, or evaluate the game state from the team locations.

Navigation has been addressed within the SocRob project under two different situations:

• Without the ball - This involved
  • The development of a self-localization algorithm, based on an omni-directional catadioptric vision system. The algorithm determines the robot posture from one image frame acquired by the catadioptric system, based on the correlation, in Hough Transform space, of extracted information on the field geometry with an a priori field geometric model (see picture). The posture is available for either guiding the robot (see below) or exchange of posture information with teammates. This algorithm is described in our RoboCup 2000 Workshop scientific award challenge winning paper "A Localization Method for a Soccer Robot Using a Vision-Based Omni-Directional Sensor", by Carlos Marques and Pedro Lima. There is also a video (AVI) to download illustrating this work. Notice that the robot self-localizes when it stops (sometimes does it more than once because it detects a failure) and, even after being "kidnapped" to a completely different posture, recovers and can get back to its home position (on the left of the image, near the wall).
  • The implementation of guidance algorithms with obstacle avoidance capabilities, to take the robot from its current posture (as obtained from its odometry, periodically reset by the self-localization algorithm) to a desired posture (e.g., robot home position, facing the goal with the ball in between, covering own goal between it and the ball). This is currently done when the robot does not have ball posession, using the Freezone algorithm, described in the paper " Multi-Sensor Navigation for Soccer Robots". There is also a video (AVI) to download illustrating this work. The experiments include moving towards the ball avoiding obstacles with periodic self-localization (reseting the odometry) which is noticeable by the fact that, when the robot does not see the ball, it gets back to its home position (on the left of the image, near the wall).
  
  An overall view of the whole navigation algorithm is provided by the IEEE Robotics and Automation Magazine (Vol. 11, No. 3, 2004) Multi-Sensor Navigation for Non-Holonomic Robots in Cluttered Environments paper.
• With the ball - the work on navigation without the ball was extended to the situation where a robot does have the ball (i.e., a dribble is required if other robots appear in the middle), by using a modified version of the potential fields method, which includes a special potential field (see diagram), to take into account the non-holonomicity of our robots, as well as constraints on linear and angular speeds, to avoid losing the ball, given the strong limitations on the robot "fingers" length (max one third of the ball diameter). The work is described in the paper A Modified Potencial Fields Method for Robot Navigation Applied to Dribbling in Robotic Soccer , and there is also a video to download illustrating this work.

Geometric "absolute" self-localization (i.e., posture coordinates w.r.t. a reference coordinate system) is surely helpful, but trusting it too much in the presence of uncertainties which sometimes cause wrong posture estimates may be dangerous, especially if many behaviors depend on it. We are currently investigating methods of "relative" navigation (e.g., moving close to a goal, or going around an obstacle) to combine them with our current solution.

More recent work has focused on MCL using the field lines + gyrodometry (check publications page), as well as on motion control algorithms taking advantage of the robots omnidirectional characteristic.