Grown-up NimbRo Robots Winning RoboCup 2017 Humanoid AdultSize Soccer Competitions
The ongoing evolution of the RoboCup Humanoid League led in 2017 to the introduction of one vs. one soccer games for the AdultSize robots, which motived our team NimbRo to enter this category. In this paper, we present the mechatronic design of our upgraded robot Copedo and the newly developed NimbRo-OP2, which received the RoboCup Design Award. We also describe improved approaches to visual perception of the game situation, including compassless localization on a soccer field with symmetric appearance, and the generation of soccer behaviors. At RoboCup 2017 in Nagoya, our robots played very well, winning the AdultSize soccer tournament with high scores. Our robots also won the technical challenges and we present the developed solutions.
With the ongoing evolution of its rules, RoboCup Humanoid League, competition each year is becoming more realistic and closer to playing games against human soccer players. In 2017, for the first time, AdultSize class robots were playing one vs. one matches, which meant bringing the soccer game to a larger scale, along with all of its inherent difficulties. In previous years, our team NimbRo has competed in the Humanoid League within the KidSize, and more recently, the TeenSize classes [1, 2, 3, 4].
In order to compete in the AdultSize category, it was necessary to procure hardware that is not only in compliance with the new set of rules, but also performs well under the conditions that a robot may face during the competition. For that purpose, we have prepared two robots capable of play: Copedo and NimbRo-OP2. Our robust hardware played a major role in our success, where over the span of five games in the soccer tournament and a set of technical challenges, our robots secured victory in both, without the need for repairs or maintenance. We describe the mechanic design of our robots in Section 2, and our approach for the technical challenges in Section 5.
Our open-source software is advanced with each RoboCup, with many modules being added and existing ones being improved. With the introduction of identically looking field halves in 2015 and the ban on magnetometers in 2017, localisation has become much more difficult. The approach we used for breaking the symmetry and localising in these conditions, as well as how we combine this information with our ball position estimate to plan further actions are described in Section 3.
2 Mechanic Robot Design
Copedo was constructed to compete at RoboCup 2012 in Mexico and played in the TeenSize class until 2015. Initially, Copedo was 114 cm tall and had a weight of 8 kg . To comply with the AdultSize class rules, our robot was extended to be 131 cm tall and 10.1 kg in weight. The upgrade of Copedo was based on another one of our robots—Dynaped, as it was upgraded in 2016 . Copedo has been fitted with new electronics and onboard PC to use our open source ROS-based framework just like Dynaped, and our other robots. Copedo is constructed from milled carbon fiber parts that are assembled to rectangular shaped legs and flat arms. The torso is constructed entirely from aluminum and consists of a cylindrical tube and a rectangular cage that holds the main electronic components. During the upgrade process, the cylindrical tube was replaced with a longer one, while the cage was that of the NimbRo-OP. To comply with the vision system of our other robots, Copedo received a new 3D printed head, similar to the one used in the igus® Humanoid Open Platform . The process and result of the hardware modifications can be seen in Fig. 2.
With a new head allowing for pan and tilt motions, Copedo has 14 actuated degrees of freedom (DoF). The hip roll, hip pitch, and knee DoF are driven by master-slave pairs of Robotis Dynamixel EX-106+ actuators. All other DoFs are driven by single actuators including EX-106+ actuators for ankle roll, EX-106+ actuators for hip yaw and RX-64 actuators for shoulder pitch, as well as the neck yaw and pitch. The robot has been fitted with cleats in the corners of its feet, to assist walking on artificial grass. More details on the robot’s core mechanical design features can be found in .
Along with the necessary structural upgrade, Copedo has received a new Intel NUC computer with an Intel Core i7-7567U processor operating at a maximum frequency of . The PC is fitted with of RAM and a solid state disk. Available communication interfaces include USB 3.0, HDMI, Mini DisplayPort and Gigabit Ethernet. The PC is connected to a Robotis CM740 board, which communicates with all actuators on a RS485 star topology bus. The CM740 incorporates a 3-axis accelerometer and gyroscope for a total of 6 axes of inertial sensory data.
According to the requirements of the RoboCup 2017 Humanoid league, we developed a new platform: the NimbRo-OP2 . Being almost 135 cm tall and only 18 kg in weight, the robot is able to participate in both TeenSize and AdultSize classes. The robot’s exoskeleton was 3D printed with an industrial-grade SLS printer out of PA-12 nylon, which contributes greatly to its low weight. In terms of electronics, the robot uses the same Intel NUC in combination with the Robotis CM740 as Copedo. There also exists the possibility of further upgrading the unit, through the inclusion of standard VESA75 and VESA100 mounts.
The structure of the robot has been simplified as much as possible to maintain low complexity, but retain functionality that is required when playing soccer. The robot in its entirety uses 34 Dynamixel MX-106 actuators. The upper body kinematics consists of three serial chains—two arms and a neck connecting the head to the trunk. The neck consists of a yaw and a pitch joint, while the arms have two pitch and one roll joint. In the legs, we decided to use parallel kinematics along with external gearing to allow for more torque output, which is necessary for dynamic walking.
The design made use of our long experience with building humanoid robots. Creating a new robot is a complex and time-consuming process, of which manufacturing plays a big role. By taking advantage of the versatility of 3D printing, we were able to develop an affordable, customisable, highly-capable, adult-sized humanoid robot in little time. The whole design and manufacturing process took less than six months. A view of the finished robot and some of its key design features can be seen in Fig. 3.
Simplicity was one of the key points in creating the design. This is best shown when looking at the knee of the robot, which houses eight actuators with only three plastic parts, two of which are covers that are not critical for bearing the load. Upon removal of the covers, servos are easily accessible for any maintenance that might be needed. The servos themselves are exposed through strategic venting hole placement to help with cooling. Most of the parts in the design are symmetrical, meaning that a single part is used in multiple spots. This minimises the amount of spare parts needed for repairs, in case a part should break. This can be seen in the ankle and hip joints, the knee joints as well as the thigh and shank links.
3 Software Design
3.1 Visual Perception
Our humanoid robots perceive the environment through digital cameras. Each robot is equipped with one Logitech C905 camera, utilising a wide-angle lens with an infrared cut-off filter. In this configuration, its field-of-view is nearly 150. Our vision system, amongst many other features, can reliably perceive game-relevant objects using texture, shape, brightness, and colour information. We project each object into egocentric world coordinates by using the intrinsic and extrinsic camera parameters. Variations in hardware result in projection errors, which scale with distance from the object. We calibrate the position and orientation of the camera frame with the Nelder-Mead  method. More details on our vision system can be found in .
A number of landmarks can be distinguished on the field, which can be used for localisation. These include line junctions, goal posts, penalty marks and the center circle. Field lines are the most useful source of information when it comes to localisation. We detect them by using a Canny edge detector on the V channel in the HSV colour space. We then apply a probabilistic Hough line detector  to extract line segments of minimum size to filter out white non-line objects. These segments are then connected to produce less, but longer lines, finally achieving a set of lines that correspond to field lines and the center circle. An example output of our vision system detecting selected objects can be seen in Fig. 4.
Localisation and Breaking the Symmetry:
To solve the global localisation problem, our method relies on having a source of global yaw rotation of the robot. We used to utilise the compass sensor in previous years, but due to the magnetometer ban, introduced in the rules of the Humanoid League in 2017, we now use the integrated gyro measurements as the source for yaw orientation. In our case, gyro integration is a reliable source of orientation tracking, but it needs a global reference. In order to set the initial heading, we could either use manual initialisation or automatic initial orientation estimation. Although manual initialisation can be done once before the start of each game, it can fail during the match. Sometimes restarting the operating system of the robot is unavoidable, which will force a reinitialisation of the heading. As a result, we reformulated the heading initialisation problem as a classification task.
According to the rules, there are a few predefined positions and orientations that the robot can start in or enter the game from. As shown in Fig. 5, the robot can start in four different positions. In two of the spots, it should face the opponent goal—near the center circle and goal area. The other two sets of locations are at the sideline in the robot’s own half—facing the field. We employ a multi-hypothesis version of our localisation module, which is initialised with four instances of initial hypothetical locations. During a brief period at the beginning, the robot tries to find the most probable hypothesis among all running instances. This stops when either the process times out or the robot finds the best hypothesis. Finally, the vision module keeps the best instance and discards the rest. To make sure that the decision is correct, we double check the result based on the perceived landmarks like goalposts and the center circle.
Obstacle detection is a crucial ability in the game, especially when the robot is handling the ball. In our software, obstacle detection is done mainly based on a model of colour distribution on the perceived robots. By having the minimum and maximum height of the robot in each size class, we roughly know what size to expect in each distance from the observer. We search for the respected bounding box size in each distance level and discard obstacle candidates that are not in the expected size range. After detecting each obstacle bounding box, we compare the colour histogram of each of the bounding boxes to the expected model of the obstacle, which are then labeled as either teammate, rival robot, or the referee. The detection history is then clustered in egocentric world coordinates and filtered based on the location of each cluster. Finally, to make the output robust against false negatives, we predict the expected movement of the obstacle in accordance to the estimated changes in the robot’s location. Each of the clusters has a certainty level, which is increased when it is detected, and decreased otherwise. The effect of this method can be seen in Fig. 4.
3.2 Soccer Behaviours
Once the visual perception module has established the current state of the game, including the localised pose of the robot on the field and the positions of the ball and any obstacles, the behaviour module uses this information to control the soccer actions of the robot. Two finite state machines (FSM) have been implemented in a hierarchical fashion on top of one another, to process the available data and output the required walking velocities and motion triggers for the robot. The higher-level FSM, called the Game FSM, is responsible for deciding on game-level actions, such as whether to try to score a goal, or perform auto-positioning, or defend the goals. The lower-level FSM on the other hand, called the Behaviour FSM, implements fundamental soccer skills such as dribbling and kicking the ball, and walking to a localised ball position while avoiding obstacles.
Approaching the ball is a fundamental skill on the level of the Behaviour FSM. The robot should walk as efficiently as possible to the required position behind the ball, without pushing the ball away. A circular halo is established around the ball, and the approach phase is divided into the near and far cases. In the far case, the robot only turns and walks forward, to maximise its speed and stability in covering ground, while in the near case, slower side-stepping is added to allow the robot to get around the ball efficiently. This is illustrated on the top left in Fig. 6.
If an obstacle blocks the approach to the ball, obstacle avoidance is applied on the level of the Behaviour FSM. The robot is slowed down, yaw rotation is added to turn it away from the obstacle, and the linear velocity of the robot is rotated to limit the radial component of the velocity towards the obstacle (see top right in Fig. 6). The maximum allowed radial velocity is a function of the proximity of the obstacle, and can also be negative, pushing the robot away from the obstacle.
Obstacle Ball Handling:
If an obstacle is close to the ball or blocks the path from the ball to the goals, obstacle ball handling is applied on the level of the Game FSM. This rotates the ball target away from obstructions, and if the ball target is no longer safely in goals, or the obstacle is too close to the ball, dribbling is enforced for safety, as shown on the bottom left in Fig. 6. The correct foot to dribble with is also enforced to avoid collisions with the obstacle as much as possible.
Kicking and Dribbling:
When the robot is behind the ball and aligned with the ball target, either the Kick or Dribble behaviours activate as required, and drive the ball towards the target. When Dribble activates (see bottom right in Fig. 6), a large variance of ball position is tolerated that allows dribbling to continue. The robot however consistently tries to correct for any misalignments in its dribble approach. If the dribble lock is lost, the ball approach is started once more.
4 Performance in Soccer Tournament
At RoboCup 2017, eight teams participated in the soccer tournament. Our robot NimbRo-OP2 played three round robin games vs. teams ZSTT (Taiwan & Korea), KIS (Japan), and CIT Brains (Japan). The robot scored very reliably, such that the three games were clearly won with a total score of 26:0. In semi-final, our robots NimbRo-OP2 and Copedo played vs. team IRC (IRC) and clearly won with a score of 9:0.
In the final, our robot NimbRo-OP2 faced the robot Sweaty (Germany). Sweaty had shown good walking and kicking capabilities during the competition. The two robots, fighting for the ball are shown in Fig. 7. Our robot had clear advantages in such fights, avoiding the opponent and maintaining balance while Sweaty often lost balance and had to leave the field. NimbRo-OP2 scored reliably. After 7:0 at half time, the game ended early with a score of 11:1.
During the competition games, NimbRo-OP2 has shown a very stable and fast walk. It never lost its balance when walking in free space. The only occasions which would lead to the fall were situations when a strong collision with other robots occurred. The kicking was robust and strong, allowing us to score far goals. Finally, the localisation worked reliably during all games and obstacle avoidance allowed to win fights for the ball by dribbling it around the opponent, which was often the case in the finals.
5 Technical Challenges
In addition to the main competition, technical challenges test how well a robot can perform a specific task in isolation. At RoboCup 2017, four technical challenges were posed: Push Recovery, High Jump, High Kick, and Kick a Moving Ball. In this section we present our strategy for three technical challenges which were addressed for RoboCup 2017.
5.1 Push Recovery
The goal of this challenge is to withstand a strong push. An impact is applied to the robot on the level of the CoM by a pendulum. To define the impulse, the 3 kg weight is retracted by a distance from the point of contact with the robot. The push is applied only from the front and from the back. The robot has to be walking on the spot during the whole challenge.
5.2 High Jump
The goal of high jump is to jump as high as possible and remain airborne as long as possible. The challenge was performed using a predesigned jump motion, which was constructed with a keyframe editor. Execution of this motion on Copedo resulted in a partially successful trial. The robot jumped higher than 10 cm, but upon landing a stable posture was not held. Copedo performing the jump is depicted in Fig. 8.
5.3 Goal Kick from Moving Ball
This challenge requires a robot to score a goal by kicking a moving ball into the goal. The ball is rolled from a ramp along the goal area line. The difficulty is defined by the angle of the ramp, and the distance from the endpoint of the ramp to the ball in its initial static position. We solved for this task is as follows:
Before the ball is released, execute a pre-kick motion. During this motion, the robot goes from a standard standing posture to a posture with one support leg. The other leg is lifted and folded in order to be ready for kicking.
Estimate the time needed for the ball to arrive in the region where it can be kicked into the goal by the previously folded leg.
Execute the kicking motion when the ball is in the kick region according to the estimation from the previous step.
The motions described above are designed using our keyframe editor. In order to estimate the time needed for the ball to arrive in the kick region, we utilise the ball detection information obtained from the vision module, described in Section 3.1. Ball observations arrive in real-time and are stored in a stack , where , where is a confidence of the detection, is the estimated position of the ball, and is a time stamp of the measurement. The position is represented in a 2D coordinate frame which covers the surface of the playing field and has its origin at the position of the robot.
We estimate the velocity of the ball with the two latest measurements , where is a predefined minimal confidence for the measurement, and is a predefined minimal time difference between two measurements. Given such measurements and , we estimate the velocity :
where is the Euclidean distance. Each obtained estimate is pushed into a double-ended queue of maximum size . We smooth the velocity by averaging over :
Finally, we estimate the time of arrival of the ball using the smoothed velocity and the most recent pair of measurements:
where is a predefined position of the ball which is optimal for performing the kick.
This simple strategy worked well for this challenge, because the ball moves with a velocity that is close to uniform. In addition, the region where the ball can be kicked into the goal is by far not limited to the position . That is why small deviations of the estimate from the ground truth do not cause failures. At RoboCup 2017, we used the following parameters: , and . NimbRo-OP2 was able to score a goal from a moving ball with the largest possible , as shown in Fig. 9.
In this paper, we described the mechatronic design of our robots and some of the perception and control
approaches that led to our success in the Humanoid League AdultSize soccer competition and
technical challenges. At RoboCup 2017, we participated in the AdultSize category for the first time
and with very little time to prepare, we were able to produce two robots that were able to secure
a victory in both competitions. Over a span of five soccer games we aggregated a total score of 46:1, conceding
a single goal in the finals, and gathered 21 points from three technical challenges. A video showing the competition highlights is available online
This work was partially funded by grant BE 2556/13 of the German Research Foundation (DFG).
- D. D. Lee, S. Yi, S. G. McGill, Y. Zhang, S. Behnke, M. Missura, H. Schulz, D. W. Hong, J. Han, and M. A. Hopkins, “RoboCup 2011 Humanoid League winners,” in RoboCup 2011: Robot Soccer World Cup XV, pp. 37–50, Springer, 2011.
- M. Missura, C. Münstermann, M. Mauelshagen, M. Schreiber, and S. Behnke, “RoboCup 2012 Best Humanoid Award Winner NimbRo TeenSize,” in RoboCup 2012: Robot Soccer World Cup XVI, pp. 89–93, Springer, 2013.
- M. Missura, C. Münstermann, P. Allgeuer, M. Schwarz, J. Pastrana, S. Schüller, M. Schreiber, and S. Behnke, “Learning to improve capture steps for disturbance rejection in humanoid soccer,” in RoboCup 2013: Robot World Cup XVII, pp. 56–67, Springer, 2013.
- H. Farazi, P. Allgeuer, G. Ficht, A. Brandenburger, D. Pavlichenko, M. Schreiber, and S. Behnke, RoboCup 2016 Humanoid TeenSize Winner NimbRo: Robust Visual Perception and Soccer Behaviors, pp. 478–490. Springer, 2017.
- P. Allgeuer, H. Farazi, M. Schreiber, and S. Behnke, “Child-sized 3D Printed igus Humanoid Open Platform,” in Proceedings of 15th IEEE-RAS Int. Conference on Humanoid Robots (Humanoids), 2015.
- G. Ficht, P. Allgeuer, H. Farazi, and S. Behnke, “NimbRo-OP2: Grown-up 3D Printed Open Humanoid Platform for Research,” in Proceedings of 17th IEEE-RAS Int. Conference on Humanoid Robots (Humanoids), 2017.
- J. A. Nelder and R. Mead, “A simplex method for function minimization,” The Computer Journal, vol. 7, no. 4, pp. 308–313, 1965.
- H. Farazi, P. Allgeuer, and S. Behnke, “A Monocular Vision System for Playing Soccer in Low Color Information Environments,” in 10th Workshop on Humanoid Soccer Robots, IEEE-RAS Int. Conference on Humanoid Robots, 2015.
- J. Matas, C. Galambos, and J. Kittler, “Robust detection of lines using the progressive probabilistic Hough transform,” Vision and Image Understanding, 2000.
- P. Allgeuer and S. Behnke, “Omnidirectional Bipedal Walking with Direct Fused Angle Feedback Mechanisms,” in Proceedings of 16th IEEE-RAS Int. Conference on Humanoid Robots (Humanoids), (Cancún, Mexico), 2016.