CMDragons 2006 Team Description

Transcription

1 CMDragons 2006 Team Description James Bruce, Stefan Zickler, Mike Licitra, and Manuela Veloso Carnegie Mellon University Pittsburgh, Pennsylvania, USA Abstract. This paper gives an overview of the team CMDragons 2006, as well as a description of some on the significant research represented in this year s team. The overview describes both the robot hardware and the overall software architecture of our team. Significant technical improvements are represented in sections on one-touch ball control and automatic tuning of the acceleration driving parameter. 1 Introduction Our RoboCup Small-Size League entry, CMDragons 2006 builds on the successful joint team CMRoboDragons, which entered in RoboCup 2004 and 2005 and placed fourth overall both years. It also builds on the prior experience of CM- Dragons entries in and [1 3]. Our team entry consists of five omni-directional robots controlled by an offboard computer. Sensing is provided by two overhead mounted cameras linked to framegrabbers on the offboard computer. The software then sends driving commands to the individual robots. Each robot has four drive wheels, a kicker, and a dribbler. The robot can also send status data back to the offboard control computer to augment the overhead vision information. The following sections describe the robot hardware and the offboard control software required to implement a robot soccer team. 2 System Overview 2.1 Robot Hardware Our team consists of five homogeneous robot agents. Each robot is omni-directional, with four custom-built wheels driven by 30 watt brushless motors. Each motor has a reflective quadrature encoder for accurate wheel travel and speed estimation. The kicker is a large diameter custom wound solenoid attached directly to a kicking plate, which offers improved durability compared to the previous design which used a standard D-frame solenoid pushing a kicking plate with guide rods. Ball catching and handling is performed by an actuated, rubber-coated dribbling bar which is mounted on a hinged damper for improved pass reception. The robot drive system and kicker are shown in Figure 1. Our robot fits within the maximum dimensions specified in the official rules, with a maximum

2 Fig. 1. Top view of the robot drive system, kicker, and dribbler. diameter of 178mm and a height of 143mm. The dribbler holds up to 19% of the ball when receiving a pass, and somewhat less when the ball is at rest or during normal dribbling. The robot electronics consists of an ARM7 core running at 58MHz linked to a Xilinx Spartan2 FPGA. The ARM7 core handles communication, runs the PD control calculations, and monitors onboard systems. The FPGA implements the quadrature decoders, PWM generation, serial communication with other onboard devices, and operates a beeper for audible debugging output. The main electronics board, which integrates all electronic components except for a separate kicker board and IR ball sensors, is shown in Figure 2. This high level of integration helps to keep the electronics compact and robust, and helps to maintain a low center of gravity compared to multi-board designs. Despite the small size, a reasonable amount of onboard computation is possible. Specifically, by offloading the many resource intensive operations onto the FPGA, the ARM CPU is freed to perform relatively complex calculations. For improved angular control, the robot also incorporates an angular rate gyroscope. It is linked into the drive control system via a secondary proportional controller, and can be used to move to or maintain a specific heading. Whenever the robot is visible to the overhead camera, the coordinate system is updated periodically using the angle determined by the vision system. This allows the local heading on the robot to match the heading for the global coordinate system.

3 Fig. 2. The main robot electronics board for CMDragons Using the gyro and secondary control loop, the robot is able to maintain a stable heading even during periods of no vision lasting several seconds. 2.2 Software The software architecture for our offboard control system is shown in Figure 3. The major organizational components of the system are a server program which performs vision and manages communication with the robots. Two other major client programs connect to the server via UDP sockets. The first is a soccer program, which implements the soccer playing strategy and robot navigation and control, and the second is a graphical interface program for monitoring and controlling the system. The server program consists of vision, tracker, radio, and a multi-client server. The vision system uses CMVision2 for low-level image segmentation and connected region analysis [4, 5]. On top of this system lies a high-level vision system for detecting the ball and robot patterns. Our robot pattern detector uses an efficient and accurate algorithm for multi-dot patterns described in [6]. Tracking is achieved using a probabilistic method based on Extended Kalman-Bucy filters to obtain filtered estimates of ball and robot positions. Additionally, the filters provide velocity estimates for all tracked objects. Further details on tracking are provided in [2]. The radio system sends short commands to each robot over a RS232 radio link. The system allows multiple priority levels so that different clients may control the same robots with appropriate overriding. This allows, for example, a joystick tele-operation program to override one of the soccer agents temporarily while the soccer system is running.

4 Camera Camera World Model Strategy Playbook Vision Tactics Radio Tracker Navigation Viewer Server Logging Server Soccer GUI Fig. 3. The general architecture of the CMDragons offboard control software. The soccer program is based on the STP framework [2]. A world model interprets the incoming tracking state to extract useful high level features (such as ball possession information), and act as a running database of the last several seconds of overall state history. This allows the remainder of the soccer system to access current state, and query recent past state as well as predictions of future state through the Kalman filter. The highest level of our soccer behavior system is a strategy layer that selects among a set of plays [7, 1]. Below this we use a tree of tactics to implement the various roles (attacker, goalie, defender), which in turn build on sub-tactics known as skills [2]. One primitive skill used by almost all behaviors is the navigation module, which uses the RRT-based ERRT randomized path planner [8 10] combined with a dynamics-aware safety method to ensure safe navigation when desired [11]. It is an extension of the Dynamic Window method [12, 13]. The robot motion control uses trapezoidal velocity profiles (bang-bang acceleration) as described in [1, 2]. 3 Significant Developments 3.1 One-touch Ball Control One significant contribution developed shortly before the 2005 competition was CMRoboDragons ability to accurately redirect a moving ball. The system we have developed combines our existing ball interception and target selection routines with a method for determining the proper angle to aim the robot to accurately redirect the ball to the target. In order to kick an incoming ball in a different direction without explicitly receiving the ball, the most important necessary new component was a model of how interaction with the kicker will affect the speed of the ball. In particular, while the kicker adds a large forward compo-

5 Fig. 4. System model for one-touch ball control. The final velocity of the ball v 1 still contains a component of the initial velocity v 0, rather than running parallel to the robot heading R h. nent to the ball velocity, effects from the ball s original (incoming) velocity are still present and non-negligible. After measurement and testing of several models, we ended up using a simple linear damping model. The system model is shown in Figure 4. The initial (incoming) velocity of the ball is denoted as v 0, while the final velocity after kicking is denoted v 1. Normalized heading and perpendicular vectors for the robot are R h and R p, respectively. The kicker provides an impulse to the ball in the direction of R h, propelling a ball initially at rest to speed k (i.e. v 0 = 0 v 1 = k). Using this model, we can estimate the final velocity using the following equation: ˆv 1 = kr h + β(r p v 1 )R p (1) In our testing, we found values of β ranging from 0.8 with no dribbler present, down to 0.3 for a robot with a dribbler spinning at maximum speed. Of course, simply having a forward model is not sufficient, as the control problem requires solving for the robot angle given some relative target vector g. However, it is easy to calculate this using a bisection search. The bounding angles for the search are the original incoming ball angle (where the residual velocity component would be zero) and the angle of target vector g (where we would aim for an ideal kicker where β = 0. The actual solution lies somewhere in between, and we can calculate an error metric e by setting up the equation above as a function of the robot angle α. R h (α) = cos α, sin α R p (α) = sin α, cos α ˆv 1 (α) = kr h (α) + β(r p (α) v 1 )R p (α) e(α) = ˆv 1 (α) g

6 Thus when e(α) > 0 the solution lies with α closer to g, while if e(α) < 0 the solution is closer to v 0. A solution at the value of α where e(α) = 0, so bisection search is simply terminated whenever e(α) < ɛ. While it is possible to invert many models so that search is not required, using a numerical method for determining α allowed rapid testing of different models, since only the forward calculation needed to be made. Bisection search has proven quite fast in practice since the calculations for a forward model are relatively simple, and only a logarithmic number of evaluations need to be made to achieve the desired accuracy. Overall, the one touch ball control proved quite useful, allowing our team last year to score most of its goals via passing, even with a relatively slow kicker 3.75m/s. This year we have applied the system to a continuous passing approach to the passing challenge, and results have so far been promising. 3.2 Automatic Tuning of Motion Parameters 2 acc=3.5, dec= distance to target (in m) time elapsed (in s) Fig. 5. The dashed curve indicates the desired theoretical execution of the motion profile; the solid curve is the actual observed motion. The left vertical bar marks the desired point in time where the target should have been reached and the robot should have completely stopped. The right vertical bar marks the observed point in time where the robot has reached the target and come to a full stop.

7 2.5 acc=3.5 dec= distance to target (in m) time elapsed (in s) Fig. 6. The meaning of the curves are similar to Figure 5. However, we can notice that both the amplitude and the total time of overshooting have been significantly reduced. One novel feature of our entry this year is the automatic optimization of motion profile parameters. Our robots are controlled according to simple trapezoidal motion profiles which are defined by an acceleration constant, a deceleration constant and a maximum speed. Traditionally, these constants needed to be hand-tweaked which was not only time intensive, but also often delivered very suboptimal and domain-dependent results (e.g. due to surface differences on different soccer fields). We overcome this problem by introducing an automated robot behavior which performs a search for the optimal set of motion profile parameters. In this automatic search routine, the robot performs several straight, equally long distance runs across the field. On each run it tries out a different set of motion parameters and evaluates their quality. The metric of motion optimality is defined as a value function which rewards speed, but punishes motion instability. The most easily detectable symptom of motion instability is overshooting of a target point which can be determined by our vision system with fairly high accuracy. One such example of overshooting can be seen in Figure 5, where the deceleration constant was too high for the robot to come to a stable halt at the target point. Instead, the robot slid through the target and had to back up which cost valuable time and precision. For comparison, we can look at Figure 6

8 where a lower deceleration constant resulted in far less overshooting, but also decreased overall speed. Our algorithm attempts to balance this tradeoff between motion instability and overall robot velocity. When put within reasonable bounds, the algorithm finishes execution after approximately five minutes which is short enough to be run before games at actual RoboCup competitions. In the near future we might attempt to extend this approach by using on-board sensors to gain more accurate stability metrics. 4 Conclusion This paper gave a brief overview of CMDragons 2006, covering both the robot hardware and the overall software architecture of the offboard control system. The hardware has built on the collective experience of our team and continues to advance in ability. The software uses our proven system architecture with many improvements to the individual modules. This paper also gave two concrete technical improvements our team has made and expects to apply this year. These include the one-touch ball control as well as a method for automatic optimization of drive parameters. We look forward to continued research on the way to this year s competition. References 1. Bruce, J., Bowling, M., Browning, B., Veloso, M.: Multi-robot team response to a multi-robot opponent team. In: Proceedings of the IEEE International Conference on Robotics and Automation, Taiwan (2003) 2. Browning, B., Bruce, J., Bowling, M., Veloso, M.: Stp: Skills tactics and plans for multi-robot control in adversarial environments. In: Journal of System and Control Engineering. (2005) 3. Murakami, K., Hibino, S., Kodama, Y., Iida, T., Kato, K., Naruse, T.: Cooperative soccer play by real small-size robots. In: Proceedings of the 2003 RoboCup Symposium. (2003) 4. Bruce, J., Balch, T., Veloso, M.: Fast color image segmentation for interactive robots. In: Proceedings of the IEEE Conference on Intelligent Robots and Systems (IROS), Japan (2000) 5. Bruce, J.: CMVision realtime color vision system. (The CORAL Group s Color Machine Vision Project) jbruce/cmvision/. 6. Bruce, J., Veloso, M.: Fast and accurate vision-based pattern detection and identification. In: Proceedings of the IEEE International Conference on Robotics and Automation, Taiwan (2003) 7. Bowling, M., Browning, B., Veloso, M.: Plays as effective multiagent plans enabling opponent-adaptive play selection. In: Proceedings of International Conference on Automated Planning and Scheduling (ICAPS 04). (2004) 8. Bruce, J., Veloso, M.: Real-time randomized path planning for robot navigation. In: Proceedings of the IEEE Conference on Intelligent Robots and Systems (IROS). (2002) 9. LaValle, S.M., James J. Kuffner, J.: Randomized kinodynamic planning. In: International Journal of Robotics Research, Vol. 20, No. 5. (2001)

9 10. James J. Kuffner, J., LaValle, S.M.: Rrt-connect: An efficient approach to singlequery path planning. In: Proceedings of the IEEE International Conference on Robotics and Automation. (2000) 11. Bruce, J., Veloso, M.: Safe multi-robot navigation within dynamics constraints. Proceedings of the IEEE, Special Issue on Multi-Robot Systems (2006, to appear) 12. Fox, D., Burgard, W., Thrun, S.: The dynamic window approach to collision avoidance. IEEE Robotics and Automation Magazine 4 (1997) 13. Brock, O., Khatib, O.: High-speed navigation using the global dynamic window approach. In: Proceedings of the IEEE International Conference on Robotics and Automation. (1999)