============== Page 1/1 ============== Computer Vision and Control for Autonomous Robots Prof. Dr. Raul Rojas FU Berlin Embodied Intelligence: A new Paradigm for AI - Intelligence needs a body: mechanics - Computer vision in real time Intelligence is the art and science -„Artificial Energy management the subconscious “ -ofLocal control - Communication between agents - Coordination and team behavior - Adaptation and learning Robotic Soccer as AI Benchmark RoboCup started with IJCAI 1997 „ I - Simulation league „ II – Small size league „ III- Mid-size league „ IV- Legged league „ V – Humanoid league „ Small-Size Liga 18 cm in diameter 4.5 by 5 meter field Five vs five Lisbon 2004 Kicking the distance Mid-size league four on four field 12 × 8 meters Lisbon 2004 Pressuring the goalie Our small-size robots Omnidirectional Design Omnidirectional Control Our mid-size robots Omnidirectional vision - Laptop for control - Firewire video camera CAD Design FUXABOT: The Hexapod I Global vision Global vision computer wireless communication global camera The world is colored Team color ball Projective Transformation Automatic camera calibration Illumination artifacts Adaptive color maps Tracking helps computer vision • the position of the ball is predicted • variable search frame • just a few pixels are read Tracking the robots We need the data of the future Data from the past t vision delay communication delay Predict the robot‘s movement positions (x1,y1) (vx,vy,w)1 (x2,y2) (vx,vy,w)2 (x3,y3) (vx,vy,w)3 (x4,y4) commands (vx,vy,w)4 θ) predictor (x,y, predictor Position and orientation four frames in the future (100 ms) II Local vision Our first omnivision robots Spherical and parabolic transformations The field seen with our mirror Locating the robot Parabolic Mirror 500 distance 400 300 200 100 0 0 0.1 0.2 pixel distance 0.3 0.4 Expectation-Maximization The model „attracts“ the cloud of points Forces on real data Precomputed resultant forces for each coordinate Obstacle Detection Obstacle Modelling Obstacle Fusion Obstacle Identification III Reactive Behavior Reactive Behavior Control slow fast sensors behaviors actuators Structure of a layer Higher layer sensors effectors sensors behaviors actors Lower layer Kicking reflex Kicking reflex activated Screenshot of control software IV Learning and Coaching the robots Anpassbarkeit Raumfreiheit Beispiel-Eingabe Learning to pass Passing game Team Play The Goalie Goalie again Learning: robot heal yourself Learn what the robot does positions (x1,y1) (vx,vy,w)1 (x2,y2) (vx,vy,w)2 (x3,y3) (vx,vy,w)3 (x4,y4) commands (vx,vy,w)4 predictor predictor (x,y,θ ) Measured position Invert the prediction (x1,y1) (vx,vy,w)1 (x2,y2) (vx,vy,w)2 (x3,y3) (vx,vy,w)3 (x4,y4) (vx,vy,w)4 predictor predictor (x,y,θ ) desired position One burnt motor V Summary and Outlook FU Fighters 1999 „ 2000 „ 2001 „ 2002 „ 2003 „ „ „ „ „ 2004 Vizeweltmeister Europa- und Vizeweltmeister Vierter Platz Europa- und Vizeweltmeister Europameister Dritter Platz (small-size) Halbfinalist (mid-size) Weltmeister (small-size) Vierter Platz (mid-size) Small-Size Team Anna Egorova, Alexander Gloye, Mark Simon, Cüneyt Göktekin, Bastian Hecht, Achim Liers, Oliver Tenchio, Fabian Wiesel, Lina Ourima, Maria Jütte, Thomas Sunderman Susanne Schöttker-Söhl Mid-size Team Holger Freyther, Ketill Gunnarsson, Henning Heinold, Felix von Hundelshausen, Wolf Lindstrot, Marian Luft, Slav Petrov, Michael Schreiber, Frederik Zilly, Fabian Ruff, David Schneider, Markus Kettern Detlef Mü M ller und Feinwerktechnik Fritz-Haber-Institut Georg Heyne „ Peter Zilske „ Torsten Vetter „ Ronald Nehring „