RoboCup 2022: Robot World Cup XXV

Object recognition is one of the key tasks in robot vision. In RoboCup SPL, the Nao Robot must identify objects of interest such as the ball, field features et al. These objects are critical for the robot players to successfully play soccer games. We propose a new statistical learning method, Class Conditional Gaussian Mixture Model (ccGMM), that can be used either as an object detector or a false positive discriminator. It is able to achieve a high recall rate and a low false positive rate. The proposed model has low computational cost on a mobile robot and the learning process takes a relatively short time, so that it is suitable for real robot competition play.

The ability to correctly anticipate an opponent’s next action in real-time adversarial environments depends on both, the amount of collected observations of that agent’s behavior as well as on the capability to incorporate new knowledge into the opponent model easily. We present a novel approach to instance-based action prediction that utilizes graph-based structures for the efficiency of retrieval, that scales logarithmically with the amount of training data, and that can be used in an online and anytime manner. We apply this algorithm to the use case of predicting a dribbling agent’s next action in Soccer Simulation 2D.

The RoboCup Small Size League employs cylindrical robots of 15 cm height and 18 cm diameter. Presently, most teams utilize a Kalman predictor to forecast the trajectory of other robots for better motion planning and decision making. The predictor is limited for such task, for it typically cannot generate complex movements that take into account the future actions of a robot. In this context, we introduce an encoder-decoder sequence-to-sequence neural network that outperforms the Kalman predictor in trajectory forecasting. The network consists of a Bi-LSTM encoder, an attention module and a LSTM decoder. It can predict 15 future time steps, given 30 past measurements, or 30 time steps, given 60 past observations. The proposed model is roughly 50% more performant than a Kalman predictor in terms of average displacement error and runs in less than 2 ms. We believe that our new architecture will improve our team’s decision making and provide a better competitive advantage for all teams. We are looking forward to integrating it with our software pipeline and continuing our research by incorporating new training methods and new inputs to the model.

In this paper, we propose an approach for gait phase detection for flat and inclined surfaces that can be used for an ankle-foot orthosis and the humanoid robot Sweaty. To cover different use cases, we use a rule-based algorithm. This offers the required flexibility and real-time capability. The inputs of the algorithm are inertial measurement unit and ankle joint angle signals. We show that the gait phases with the orthosis worn by a human participant and with Sweaty are reliably recognized by the algorithm under the condition of adapted transition conditions. E.g., the specificity for human gait on flat surfaces is 92 %. For the robot Sweaty, 95 % results in fully recognized gait cycles. Furthermore, the algorithm also allows the determination of the inclination angle of the ramp. The sensors of the orthosis provide 6.9\(^\circ \) and that of the robot Sweaty 7.7\(^\circ \) when walking onto the reference ramp with slope angle 7.9\(^\circ \).

This work introduces a ultra-fast object detection method named FLA-CNN for detecting objects in a scene from a planar LIDAR signal, using convolutional Neural Networks (CNN). Compared with recent methods using CNN on 2D/3D lidar scene representation, detection is done using the raw 1D lidar distance signal instead of its projection on a 2D space, but is still using convolutional neural networks. Algorithm has been successfully tested for RoboCup scene analysis in Middle Size League, detecting goal posts, field boundary corners and other robots. Compared with state of the art techniques based on CNN such as using Yolo-V3 for analysing Lidar maps, FLA-CNN is 2000 times more efficient with a higher Average Precision (AP), leading to a computation time of \(0.025\,ms\), allowing it to be implemented in a standard CPU or Digital Signal Processor (DSP) in ultra low-power embedded systems.

This work presents a study for building a Deep Vision pipeline suitable for the Robocup Standard Platform League, a humanoid robot soccer tournament. Specifically, we focus on end-to-end trainable object detection for effective perception using Aldebaran NAO v6 robots. The implementation of such a detector poses two major challenges, those of speed, and resource-effectiveness with respect to memory and computational power. We benchmark architectures using the YOLO and SSD detection paradigms, and identify variants that are able to achieve good detection performance for ball detection, while being able to perform rapid inference. To add to the training data for these networks, we also create a dataset from logs collected by the UT Austin Villa team during previous competitions, and set up an annotation pipeline for training. We utilize the above results and training pipeline to realize a practical, multi-class object detector that enables the robot’s vision system to run 35 Hz while maintaining good detection performance.

In the RoboCup@Work competition, the Rotating Table Test problem refers to the task of automatically grasping an object from a circular table, rotating at constant angular velocity. This task requires the robot to track the target object’s position and grasp it. In this work, we propose a camera-based online tracking system which works in real-time. Our approach is based on the YOLOv5 detection backbone and uses a novel, modified version of the SORT tracker. The tracker is trained solely on a pre-existing detection dataset containing annotated static images, thanks to which the collection of additional situation-specific video data is not required. We evaluate and compare SORT with YOLOv5 and SqueezeDet backbones and demonstrate the improvement in tracking performance when using the former. The evaluation dataset and corresponding annotations are made available for use in the community.

Due to increased demands related to flexible product configurations, frequent order changes, and tight delivery windows, there is a need for flexible production using AI methods. A way of addressing this is the use of temporal planning as it provides the ability to generate plans for complex goals while considering temporal aspects such as deadlines, concurrency, and durations. A drawback in applying such methods in dynamic environments is their high and unpredictable planning time. In this paper, we present an evaluation of the current state-of-the-art temporal planners within the RoboCup Logistics League. Among the many factors that impact automated planners applicability, the level of abstraction of the planning model is paramount. We center our study on the effect that modeling choices have on the performance of the assessed planners. Our experimental results suggest that seeking for the right level of abstraction of planning domain models allows for compromising solutions between plan quality and plan solving time.

In the RoboCup Small Size League (SSL), teams are encouraged to propose solutions for executing basic soccer tasks inside the SSL field using only embedded sensing information. Thus, this work proposes an embedded monocular vision approach for detecting objects and estimating relative positions inside the soccer field. Prior knowledge from the environment is exploited by assuming objects lay on the ground, and the onboard camera has its position fixed on the robot. We implemented the proposed method on an NVIDIA Jetson Nano and employed SSD MobileNet v2 for 2D Object Detection with TensorRT optimization, detecting balls, robots, and goals with distances up to 3.5 m. Ball localization evaluation shows that the proposed solution overcomes the currently used SSL vision system for positions closer than 1 m to the onboard camera with a Root Mean Square Error of 14.37 mm. In addition, the proposed method achieves real-time performance with an average processing speed of 30 frames per second.

This paper presents the approach of the RoboCup@Work team tyrolics of the university of Innsbruck to design, develop and build a mobile manipulator with 10 degrees of freedom. The mobile manipulator LeoBot uses Mecanum wheels to enable omnidirectional movement and includes a Franka Emika Panda serial manipulator. This paper focuses on hardware development and provides information on mechanical, electronic, and mechatronic system components. Basic algorithms developed and used for the competition are briefly described.

Soccer Simulation 2D League is one of the major leagues of RoboCup competitions. In a Soccer Simulation 2D (SS2D) game, two teams of 11 players and one coach compete against each other. Several base codes have been released for the RoboCup soccer simulation 2D (RCSS2D) community that have promoted the application of multi-agent and AI algorithms in this field. In this paper, we introduce “Cyrus2D Base”, which is derived from the base code of the RCSS2D 2021 champion. We merged Gliders2D base V2.6 with the newest version of the Helios base. We applied several features of Cyrus2021 to improve the performance and capabilities of this base alongside a Data Extractor to facilitate the implementation of machine learning in the field. We have tested this base code in different teams and scenarios, and the obtained results demonstrate significant improvements in the defensive and offensive strategy of the team.

Due to the physical limitations of real robots, simulated robotics is an important area of research that opens up a lot of possibilities to study the robots’ dynamics and program their behaviors. RoboCup 3D Soccer Simulation league is a tournament to encourage the development of robots that compete in a high-fidelity simulation ambient.

Due to the high complexity of simulated humanoid robots, numerical optimization techniques are often used to determine the best parameters to control their motion sequence. Even the simplest movements, such as walking, stopping and getting up, can have a meaningful impact on a simulated soccer match if well optimized. Such a process is time consuming and has a high computational cost. For this reason, the usage of high performance clusters is a good way to accelerate the optimization, and for this technology to be used, it is necessary to develop a reliable tool that interfaces the cluster, simulation, and optimization.

The main contribution of this work is to provide a distributed optimization tool for the RoboCup 3D Soccer Simulation League based on the Intel DevCloud cluster.

This paper presents a open-source omnidirectional walk controller that provides bipedal walking for non-parallel robots through parameter optimization. The approach relies on pattern generation with quintic splines in Cartesian space. Additionally, baselines of achieved walk velocities in simulation for all robots of the Humanoid Virtual Season, as well as some commercial robot models, are provided.

An important part of ensuring continuous development and betterment of performance of a robot soccer team is the process of analyzing past matches and generating insights about possible causes for good or bad outcomes the team has in the field. RoboCup robot soccer matches from leagues such as the 2D Simulation League generate log files that can help create insights using data analysis, but since there is currently no active open collaboration between teams for building an ecosystem of analysis tools, this area is underdeveloped - as a community - and could be improved. We propose an open-source data analysis library that contains all of the basic structures needed for implementing any analysis, as well as a collection of ready-to-use, quasi-agnostic analysis that can be used to analyze matches from any soccer league with little or none adaptation. We believe this can be a common ground for developers from any team to work together in the advancement of technology and lower the barrier of entry into the data analysis realm for teams that are not yet involved in the area. We also demonstrate how this library can be leveraged as a software component for other projects, by building a custom web platform that utilizes it.

The 2D Simulation League (SIM2D) is one of the most accessible RoboCup leagues, since there are no hardware costs included, and codebases from which you can build a new team are readily available for free. The league, however, still has a notable barrier of entry: its setup. The setup process for the SIM2D environment can be daunting for newcomers, especially for people new to programming or that don’t have access to a linux based distro. In this sense, if there is interest in lowering the barrier of entry of the SIM2D league, and with it, of robotics in general, it would be helpful to have tools with minimal setup, or that don’t require any setup at all. This article reports on an Open-Source monitor for SIM2D games, “Web Soccer Monitor”, that runs entirely on the browser and doesn’t require a setup to function. It is useful in itself as it simplifies the experience of utilizing a monitor from the user’s perspective, while also providing developers with a more modern and agile framework in which to implement new features, but it also serves as the foundation in which the RoboCup community can start building an entire ecosystem of SIM2D web tools, which would lower even more the barrier of entry and would, among other things, facilitate the creation of new categories, such as a fully-fledged SIM2D Junior League.

This paper presents the approach employed by the team RoboBreizh to win the championship in the 2022 RoboCup@Home Social Standard Platform League (SSPL). RoboBreizh decided to limit itself to an entirely embedded system with no connection to the internet and external devices. This article describes the design of embedded solutions including manager, navigation, dialog and perception. We present results from the competition showing up the value of our proposal.

We describe the technologies of our autonomous soccer humanoid robot system that won the RoboCup2022 Humanoid KidSize League. For RoboCup2022, we developed both hardware and software. We developed a new hardware SUSTAINA-OP. We aimed to make it easier to build, harder to break, and easier to maintain than our previous robot. SUSTAINA-OP is an open hardware platform. As the control circuit, we selected a computer with higher processing power for deep learning. We also developed its software. In terms of image processing, the new system uses deep learning for all object detection. In addition, for the development of action decision-making, we built a system to visualize the robot’s states and solved many problems. Furthermore, kicking forward at an angle action is added as a new tactical action. In RoboCup2022, even when the robots were facing each other with the ball between them, by this action the robot succeeded in getting the ball out in the direction of the opponent’s goal.

This paper presents the team AutonOHM and their solutions to the challenges of the RoboCup@Work league. The hardware section covers the robot setup of Ohmn3, which was developed using knowledge from previous robots used by the team. Custom solution approaches for the @Work navigation, perception, and manipulation tasks are discussed in the software section, as well as a control architecture for the autonomous task completion.

Beating the human world champions by 2050 is an ambitious goal of the Humanoid League that provides a strong incentive for RoboCup teams to further improve and develop their systems. In this paper, we present upgrades of our system which enabled our team NimbRo to win the Soccer Tournament, the Drop-in Games, and the Technical Challenges in the Humanoid AdultSize League of RoboCup 2022. Strong performance in these competitions resulted in the Best Humanoid award in the Humanoid League. The mentioned upgrades include: hardware upgrade of the vision module, balanced walking with Capture Steps, and the introduction of phase-based in-walk kicks.

The RoboCup Soccer Simulation 2D Competition is the oldest of the RoboCup competitions. The 2D soccer simulator enables two teams of simulated autonomous agents to play a game of soccer with realistic rules and sophisticated game play. This paper introduces the RoboCup 2022 Soccer Simulation 2D Competition champion team, HELIOS2022, a united team from Okayama University of Science and Osaka Metropolitan University. The overview of the team’s two recent approaches is also described. The first one is the method of online search of cooperative behavior for the setplay planning. The second is a performance evaluation system for efficient team development.

This paper provides an overview of the main developments of the Tech United Eindhoven RoboCup@Home team. Tech United uses an advanced world modeling system called the Environment Descriptor. It allows for straightforward implementation of localization, navigation, exploration, object detection & recognition, object manipulation and human-robot cooperation skills based on the most recent state of the world. Other important features include object and people detection via deep learning methods, a GUI, speech recognition, natural language interpretation and a chat interface combined with a conversation engine. Recent developments that aided with obtaining the victory during RoboCup 2022 include people and pose recognition, usage of HSR’s display and a new speech recognition system.

In 2022, TIGERs Mannheim won the RoboCup Small Size League competition with individual success in the division A tournament, the blackout technical challenge and the dribbling technical challenge. The paper starts with an outline of the robot’s dribbling hardware and ball catching computations, followed by a high level summary of the AI used in the tournament. Given 62 scored goals and no conceded goals at RoboCup 2022, the focus is on describing the used attack and support behaviors and how they are selected. The paper concludes with a statistic of the tournament backing the efficiency of our employed strategies.

The B-Human team won all of its seven games at the RoboCup 2022 competition in the Standard Platform League (SPL), scoring a total of 48 goals and conceding 0. B-Human achieved this high level of performance with a new behavior architecture that enables more cooperative game play while sending fewer team communication messages. This paper presents the parts of the behavior that we consider crucial for this year’s success. We describe the strategic evaluation, action selection, and execution aspects of our new pass-oriented play style. The effectiveness of our algorithms is supported by statistics from competition games and extensive testing in our simulator. Empirically, our approach outperforms the previous behavior with a significant improvement of the average goal difference.

The RoboCup Logistics League (RCLL) is a robotics competition in a production logistics scenario in the context of a Smart Factory. In the competition, a team of three robots needs to assemble products to fulfill various orders that are requested online during the game. This year, the Carologistics team was able to win the competition with a new approach to multi-agent coordination as well as significant changes to the robot’s perception unit and a pragmatic network setup using the cellular network instead of WiFi. In this paper, we describe the major components of our approach with a focus on the changes compared to the last physical competition in 2019.

FC Portugal, a team from the universities of Porto and Aveiro, won the main competition of the 2022 RoboCup 3D Simulation League, with 17 wins, 1 tie and no losses. During the course of the competition, the team scored 84 goals while conceding only 2. FC Portugal also won the 2022 RoboCup 3D Simulation League Technical Challenge, accumulating the maximum amount of points by ending first in its both events: the Free/Scientific Challenge, and the Fat Proxy Challenge. The team presented in this year’s competition was rebuilt from the ground up since the last RoboCup. No previous code was used or adapted, with the exception of the 6D pose estimation algorithm, and the get-up behaviors, which were re-optimized. This paper describes the team’s new architecture and development approach. Key strategy elements include team coordination, role management, formation, communication, skill management and path planning. New lower-level skills were based on a deterministic analytic model and a shallow neural network that learned residual dynamics through reinforcement learning. This process, together with an overlapped learning approach, improved seamless transitions, learning time, and the behavior in terms of efficiency and stability. In comparison with the previous team, the omnidirectional walk is more stable and went from 0.70 m/s to 0.90 m/s, the long kick from 15 m to 19 m, and the new close-control dribble reaches up to 1.41 m/s.

For the first time, the HERA robot won the RoboCup@Home in the Open Platform League in Bangkok, Thailand. This robot was designed and developed by the RoboFEI@Home team, considering all mechanical, electronic, and computational aspects. It is an Open League platform capable of performing autonomous tasks in home environments, in addition to human-robot interaction, collaborating with people who share the same environment. In this edition of the competition, the platform presented advances in the methods of interacting with people and social navigation. Interaction with people and objects is supported by image segmentation processes, enhancing environment perceptions and people recognition during tasks.

During the RoboCup 2022 tournament in Bangkok, Thailand, Tech United Eindhoven achieved the first place in the Middle Size League. This paper presents the work done leading up to the tournament. It elaborates on the new swerve drive platform (winner of the technical challenge) and the progress of making the strategy software more semantic (runner-up of the scientific challenge). Additionally, the implementations of the automatic substitution and of more dynamic passes are described. These developments have led to Tech United winning the RoboCup 2022 tournament, and will hopefully lead to more successful tournaments in the future.