Artificial intelligence (AI) development has been increasing significantly in recent years, driven by advances in machine learning algorithms and hardware capabilities. While there are many different possible applications for AI, our team is particularly interested in implementing it in soccer. Sport is an important part of human culture, where we learn to test our physical limits and work as a team.
In soccer, the arrangement of 11 players on the field, known as the formation, is developed over years of playing and practicing. Formations like 4-4-2 (4 defenders, 4 midfielders, and 2 forwards) and 4-3-3 are chosen based on a team's strengths, weaknesses, and tactical objectives, as well as the strengths and weaknesses of their opponent. Deciding on the best strategy for each game is extremely challenging for coaches, as there is no formula for success. However, with machine learning, AI could teach themselves to play soccer. We believe that in future AI with machine learning configured with data on a team's and its opponents' strengths and weaknesses could provide effective statistical assistance to coaches by running thousands or even millions of simulated matches in the background.
While there is rapid growth in developing AI with different machine learning algorithms in sport. In this project we are going to dive into implementing machine learning on robot soccer playing, and simulate a real-world competitive soccer match.
The main task of the project is to design and develop an AI self-teaching robot that is able to learn to play soccer in a virtual environment, and to create a virtual environment for training and simulation that accurately replicates the rules and dynamics of a real-world soccer match. To train the AI self-teaching robot using reinforcement learning or another suitable machine learning algorithm. lastly to enable the AI self-teaching robot to work as a team with other robots to simulate a competitive soccer match in the virtual environment.
While the sub task is to identify the most effective training method by experimenting with different algorithms and settings, based on the performance of the robots in the virtual environment. and to demonstrate the capabilities of the AI self-teaching robot and the effectiveness of the training method through simulations of real-world soccer matches in the virtual environment
Our proposed solution for a robot in a soccer game is to use visual feedback and planning and control systems to enable the robot to interact with and understand its environment, additionally a virtual physical environment to allow them to perform visual soccer matches.
The visual feedback system for our robot will use image processing and classification for visual navigation, allowing the robot to avoid obstacles and stay within bounds. It will also use object detection to recognize and track objects such as the ball, teammates, and opponents, enabling the robot to perform tasks related to these objects. Additionally, we will use localization and Kalman filter to determine the position of objects and estimate the robot's location in the map over time. To enable the robot to make decisions and take actions based on the data provided by the visual feedback system, we will need to use path planning and motion control methods. To perform self teaching tasks, our proposed method is to use machine learning to train the robot to make decisions and take actions based on data, additionally reinforcement learning algorithms to adapt to changing environments and goals.
To perform a soccer match simulation, we will need to build a virtual physical environment and model. The environment should include models of a soccer pitch with two goals at each end, football and the soccer playing robot. In addition, to build the intelligent systems and behaviors for the players into the simulation, we will need an AI platform with AI-related features and tools. This platform should provide tools for developing and training the robot, as well as features for building and running simulations.
By combining the above approaches, our proposed solution will enable a robot to navigate and interact with its environment effectively in a soccer game, allowing it to perform training in simulation to achieve its goals.
Assumption
The following assumptions have been made in order to simplify the model and make it more manageable:
- The player and the ball are rigid bodies.
- The sense of vision and hearing is simulated by rays emitting from the body.
- The player and ball motion and movement is limited to forward, backward, and rotation only, and is constrained to a 2D horizontal plane.
- No offside, penalties or concern
- There are walls to limit the ball and robot to stay in the field
- There is no effect of wind or air resistance in the environment.
Additionally, we assume that the project has the necessary resources and computer power to develop the AI self-teaching robot and create the virtual environment for training and simulation.
Basic simulation platform
Throughout the design process, we have identified an open-source project called ML-Agents[1] as a suitable starting point for our soccer project, which uses reinforcement learning algorithms to achieve self-teaching soccer playing. The project utilizes the Unity game engine to create a visual simulation and the Unity ML-Agents toolkit and PyTorch library to develop the model and reinforcement learning algorithm. We believe that this existing project, which provides a basic 2v2 soccer game, is well-suited to our needs and can serve as the foundation for our goal. Our approach involves building upon this 2v2 learning environment and experimenting with our own learning algorithms, with the ultimate goal of expanding the game to a more complex 5v5 model featuring two teams of 5 robots trained by different algorithms, In addition, we plan to enhance the performance of the robots by assigning them positions such as goalie, attacker, and defender