Tech United Eindhoven Team Description 2012

Transcription

1 Tech United Eindhoven Team Description 2012 R. Hoogendijk, G.J.L. Naus, F.B.F. Schoenmakers, C.A. Lopez Martinez, G.M. Heldens, J.W.M. t Hoen, R.J.E. Merry, M.J.G. van de Molengraft Eindhoven University of Technology, Den Dolech 2, P.O. Box 513, 5600 MB Eindhoven, The Netherlands techunited@tue.nl Abstract. This paper describes the advances in the mechanical, electrical and software design of Tech United s Middle Size League robots for the past year. The main innovations include an improved performance of the goalkeeper, a more accurate position estimation, the use of la laser range finder for ball detectioncheck ball position, an improved pathplanning, and a monitoring system of the robots behavior. Keywords: RoboCup Middle Size League, goalkeeper robot, 3D ball recognition 1 Introduction Tech United Eindhoven is the RoboCup team of Eindhoven University of Technology. It consists mainly of PhD, MSc and BSc students, supplemented with academic staff members from different departments of Eindhoven Technical University. Tech United was founded in 2005 and the team is participating in the Middle Size league since Starting this year, the team will also in and the Humanoid Leagues. This paper describes the most significant improvements regarding the mechanical, electrical and software architecture of the Middle Size League robots. This document is part of the qualification package for the RoboCup MSL World Championships 2012 in Mexico City, and is organized as follows. In section 2 a brief introduction of the robot platform is presented. The main improvements compared to the previous team description [1] are indicated in section 3. Finally, we end with conclusions. 2 Robot Platform During tournaments and numerous demonstrations, the current generation soccer robots has proven to be very robust. The schematic representation published in the previous team description paper [1] still largely covers the robot, but several changes have been made to hard- and software design. Subtarget planning, for example, has been improved in a way such that scrum situations are avoided, positioning has been improved using dynamic optimization based on fuzzy rules, and the ball handling mechanism has been robustified. 2.1 Hardware The robots of Tech United Eindhoven have been named TURTLEs (acronym for Tech United RoboCup Team: Limited Edition). A picture of the fifth generation robots (2011) is shown in Fig. 1. Three 12V Maxon motors, driven by Elmec Violin 25/60 amplifiers and two Makita 24V, 3.3Ah batteries, are used to the omniwheels. The TURTLEs have an active ball handling mechanism which enables them to control the ball when driving forwards, while turning, and even when driving backwards [2]. The solenoid shooting mechanism, which is powered by a 450V, 4.7mF capacitor, provides an adjustable, accurate and powerful shot [3]. To acquire information on the surroundings, the robot has two cameras, a front camera and an omnivision camera. The high speed front camera can accurately track the ball, and is able to see the ball when it is in the air. The omnivision camera has a 360 view which gives information on the positioning and the surroundings. An electronic compass is used to differentiate between the own side of the field and the opponent side.

2 Fig. 1. Fifth generation TURTLE with cover attached (left), with cover removed (middle) and a close up of the robustified ball handling mechanism (right). 2.2 Data-Acquisition and Software To facilitate data-acquisition and motion control, the robots are equipped with EtherCAT devices [4,5], which are connected to the onboard host computer via ethernet. Each robot is equipped with a mini-pc running a pre-emptive Linux kernel (Ubuntu 8.10). The software is automatically generated from Matlab/Simulink models via the RTW toolbox. The software is divided in three main parts, namely a vision, worldmodel and motion module. While the vision and worldmodel modules both run at 30Hz, the motion module uses a much higher sampling rate of 1000Hz. The vision module provides localization of ball, TURTLEs and opponents. Hereafter the worldmodel combines information from all robots to get an unified estimation on peer and opponent positions. The motion module contains the strategy, pathplanning and the actual motion controllers of the TURTLEs. 3 Main Improvements Goalkeeper Calibration tool for FPGA platform on keeper s front camera On the keeper a front camera with FPGA platform was already installed and used last year. A problem however, was that due to the fact that it was a first implementation, the software had to be rebuild and uploaded to the FPGA every time the detection parameters for the algorithm changed. For more convenient use of the platform, a calibration tool was created. With this tool a live stream from the camera can be obtained in different views, parameters can be exchanged and the detection can be analyzed. The detection algorithm requires threshold values for HSV values and confidence levels. To easily obtain these threshold values, images can be stored and loaded into the program. One selects the ball in the image and the program returns the appropriate threshold values. These values can be fine tuned and the effect is immediately visible on the screen. In Fig. 2 a screenshot of this tool is depicted. Once these values are sent to the FPGA, the detection performance of the platform can be analyzed live on screen. Also, different views can be selected such as only a thresholded H-level image or the image after the complete HSV threshold filter, see Fig. 3 for a screenshot of this tool. An EEPROM is installed on the FPGA to save the parameter values such that they are not lost in case of a power loss.

3 Fig. 2. Calibrating the threshold for HSV values Fig. 3. The front camera calibration tool Keeper motion Last year we designed and built a new keeper robot, see Fig. 4(a). The main advantage of the new design is the four wheel configuration instead of the previous three configuration. Its four omniwheels are placed in such a way that they are almost parallel to the goal line when the robot is in the goal see Fig. 4(b). This enables high sideways accelerations. The keeper is hard to control when accerating sideways rapidly, because the wheels are slipping. This can be explained from Figure 4(b), which shows a simplified front-view of the robot. Because the wheels of the robot are located below the center of gravity (cog), the forces F 1 and F 2 generated by the wheels will not only cause a sideways acceleration a as intended, but also an angular

4 acceleration α. Because of this, the normal force n 1 will be increased while n 2 is reduced. This is why the wheels at the right side will slip if the left and right side are actuated with the same force. Therefore, the torques on the right wheels must be lower than the torques on the left side. The solution to this problem was to adapt the existing decoupling in the control software of the keeper such that the force distribution depends on the direction of the acceleration in sideways direction. In this way, when accelerating to the right, the right wheels will have lower forces. When the acceleration is to the left, the left wheels will have less force. This also provides optimal braking without slipping wheels. Before this improvement, the acceleration of the keeper was limited to 3 m/s 2. With the new decoupling, accererations up to 10 m/s 2 are possible without slip, which reduces the time to travel from the center of the goal to the goalpost from 0.85 seconds to 0.46 seconds. n 1 α cog n 2 a (a) The keeper robot of Tech United. F 1 F 2 (b) Forces on the keeper during acceleration. Fig. 4. Caption of subfigures (a) and (b) 3.2 Strategy Position estimation for self-localisation In order to broaden the game skills of the robots, it is desired to perform fast and effective interception of free balls, which enable reliable passing during game, indirect passes, etc. All these skills require an accurate estimation of the robots position, balls position and balls speed. The latter two estimations depend highly the estimate of the robots global position, which is obtained by combining measurements from vision and encoders. Vision offers a non-biased but noisy estimate of the robots global position at a rate of 32Hz. Encoders signals, which are taken at 1000Hz, exhibit low noise but are biased because the robots wheels can slip. In the current approach, the estimation is obtained from the encoders and when it deviates too much from vision measurements, the estimation is updated. The biggest problem of this method is that the sudden jumps during a reset deteriorates the estimation of the balls speed, which is of crucial importance for an appropriate interception of the ball. In order to solve this problem, the signals from vision and the encoders should be fused in a smooth way. This can be accomplished by means of a Kalman filter and a predictor, which provides an estimate of the vision data between two vision samples. Hence, at each encoder sample time,

5 the Kalman filter can be updated with (estimated) vision data. Such an estimator is successfully validated with real data from the robots, which include bias, noise and delays. The results for the x-coordinate are shown in Fig. 5. The left plot clearly shows that the estimator provides a smooth estimate of the robots position. A zoom is provided in the right plot and shows that the estimation is also robust to drops in the vision measurements. position (meters) x position (meters) x position x position vision x position (new) vision valid time (sec) time (sec) Fig. 5. Estimation of x position using different approaches Check ball possession with the laser range finder Formerly, the TURTLE determines if it has possession of the ball based on the ball position estimation from the omnivision and the position of the arms of the ball handling mechanism. False positives when checking for possession of the ball can be caused when an opponent is standing against the ball handling mechanism during a scrum. This causes the ball handling mechanism to be pushed upwards and to give a positive signal. If the ball is also in front of the robot, the error in position estimation can also give a false positive signal of ball possession. Detecting possession of the ball is crucial for the team and robot strategy. False positives during a scrum can cause the Fig. 6. Left: no ball. Middle: Ball detected. Right: False positive for ball detection.

6 robot to drive backwards to go around its opponent and attempt an attack in the direction of the opponents goal. The strategy also changes to attacking mode when one of the peers has possession of the ball. To reduce the chance of false positives, the installed Hokuyo UTM-30lx laser scanners [1] are used. The ball handling mechanism is designed such that the laser beams can pass through a hole in the ball handling arm. The holes are designed such that the laser beams are parallel to the walls of the holes to maximize the number of beams able to measure the distance to the ball, see Fig. 6. Since the radius of the ball is relatively small compared to the radius of opponent robots, the ball penetrates the ball handling mechanism much more than other objects on the field. The distance measured between the ball handling mechanism and the object in between can be used as a measure for ball possession Pathplanning The Tech United motion planning architecture consists of three layers, see Fig. 7. A strategy layer determines a global path, targeting to evade obstacles as much as possible and find the open spots on the field to dribble too. To compute the global path, a Voronoi-based algorithm is used, combining vision information of all robots [1]. The output of the strategy layer is a vector of length n>0 of target positions for the robot. The second layer prevents collisions while driving the shortest path towards the desired target position [6]. Only local vision information is used in this case. Subtargets are planned to avoid collision. The third and final layer consists of a second-order trajectory planner, taking into account velocity and acceleration constraints. The output of the third layer, is used to control the motion of the robot. shared position information local position information encoder output strategy layer global Voronoi path planning target action layer local obstacle avoidance subtarget control layer constrained trajectory planning desired acceleration Fig. 7. Tech United motion planning architecture. The second path planning layer determines the obstacles in between the desired target position and the robot position that is closest. Next, all obstacles in the neighborhood of these obstructing obstacles are clustered into a single polygonal avoidable. Based on this avoidable, a subtarget is planned such that the robot avoids a collision. An example is given in Fig. 8(a) [6]. This approach works fine as long as neither the robot nor the desired target position are enclosed by obstacles. In that case, clustering all neighboring obstacles means that either the robot or the target is positioned inside the resulting avoidable. The result is analogous to the problem of local minima that many other motion planning algorithms experience [7]. To solve this problem, an algorithm is developed that determines the best way to either escape (or enter) the local minimum (or maximum). The algorithm determines the largest open angle between two adjoining obstacles and implements a waypoint for the robot, which is illustrated in Fig. 8(b). To prevent chattering hysteresis is implemented. Currently, focus is on including obstacle velocities in the motion planning problem. This proved to be difficult in the current architecture. Therefore, the overall planning architecture is re-evaluated based on a comprehensive set of requirements. To arrive at a general set of requirements for different applications, the requirements are put together in close cooperation with team of Tech United.

7 robot avoidable obstacle target α 5 α 4 α 1 waypoint target α 3 α 2 subtarget (a) Local obstacle avoidance by subtarget (b) Waypoint planning to escape local minima. planning. Fig Monitoring To analyze the behavior of our TURTLEs more conveniently a monitoring tool with a top camera was created. A camera with a fish eye lens is installed above our testing field in Eindhoven. The video feed from this camera is then used as an underlay for our Greenfield observation tool [9]. This way the position and movement of the TURTLEs can be compared to the real world which is represented by the camera feed. In Fig. 9 the tool is displayed. Especially when analyzing the passing between TURTLEs or the path planning, this tool is very useful. The video feed and the observation data can both be saved for reviewing at any time. Fig. 9. A screenshot of the Greenfield observation tool with video underlay To correct for the deformation caused by the fisheye lens, the video feed is transformed. The video feed is exactly aligned with the Greenfield observation tool. For this transformation (and other processing), OpenCV libraries are used in the software. 4 Conclusions Compared to the situation described in the previous team description paper, Tech United Eindhoven has introduced several improvements. The improved goalkeeper results in a better defence during game play. The robots have a more accurate estimation of their position, ball position and ball speed. Also the errors for ball possesion check are reduced. The pathplanning is improved via better obstacle avoiding algorithms. To analyse the behavior of our TURTLEs a monitoring tool is created using a top view camera. All new developments together yield an improved game performance, hopefully improving last years results in the RoboCup Middle Size league.

8 5 References 1. F. Kanters, R. Hoogendijk, R. Janssen, K. Meessen, J. de Best, D. Bruijnen, G. Naus, W. Aangenent, R. van den Berg, H. van de Loo, G. Heldens, R. Vugts, G. Harkema, P. van Brakel, B. Bukkums, R. Soetens, R. Merry, M. van de Molengraft, Tech United Team Description Paper, RoboCup, J. de Best, M. van de Molengraft, M. Steinbuch, A novel ball handling mechanism for the robocup middle size league, Mechatronics 21(2). 3. K. Meessen, J. Paulides, E. Lomonova, A football kicking high speed actuator for a mobile robotic application, Proceedings of the 36th Annual Conference of the IEEE Industrial Electronics Society (2010) D. Jansen, H. Buttner, Real-time ethernet the EtherCAT solution, Computing and Control Engineering Journal 15 (1) (Feb.-March 2004) S. Potra, G. Sebestyen, EtherCAT Protocol Implementation Issues on an Embedded Linux Platform, Automation, Quality and Testing, Robotics, 2006 IEEE International Conference on 1 (May 2006) D. Bruijnen, J. van Helvoort, M. van Molengraft. (2007), Realtime motion path generation using subtargets in a rapidly changing environment, J. Robotics and Autonomous Systems, 55 (6), p J Latome. (1990), Robot motion planning, Springer 8. J. Lunenburg, T. Clephas, N. Dirkx, B. Willems, J. Elfring, J. Sandee and M. van de Molengraft, Tech United Team Description Paper, RoboCup, F. Schoenmakers, R. Janssen (2011) Greenfield Aumented Reality,