Facebook is training AI to play boxing matches. Facebook’s artificial intelligence research department is responsible for this development. The researchers associated with this project have developed a learning framework that stimulates the training route of people engaged in learning various sports to learn basic skills, round-level strategies, and much more. Moreover, they have also developed a strategy on the encoder-decoder structure that allows physical simulation roles to be trained and learned. The researchers also showcased the strategies in the framework through boxing and fencing.

One of the main challenges the researchers had to face was concerned with multiplayer animation since it requires that the interaction taking place between various characters is synchronized both in time and space. Moreover, the fact that there is very little data on the coordination of different skills makes the project much harder. In the research paper titled “Control Strategies for Physically Simulated Characters Performing Two-player Competitive Sports”, FAIR explored some of the techniques of training control systems and developed a framework that generates control strategies.

The brainchild of this program, the humanoid robot, consequently, has greater freedom and is driven by joint torque. FAIR’s project uses deep reinforcement learning for the characters to acquire basic skills and learn competition-level strategies. The framework developed by FAIR uses a set of sports data, and it, in turn, generates the control strategies of two physical simulation players which allows the characters to perform a set of basic skills using the right technique which will prompt the player to win the game.

For the simulation to work, the researchers will first have to collect some movement data and then engage the deep reinforcement learning method, and finally, the imitation strategy becomes a competitive strategy. For this, the researchers use a new strategy that consists of a task encoder and a motion decoder. However, there still are huge problems with this motion capture mainly because of the interaction among multiple agents which becomes difficult to capture. To overcome this problem, FAIR designed a framework that captures the motion data. First, they capture the action using an agent, and then the required interaction is created through simulation and learning techniques.

The model developed by FAIR needs a significant amount of calculations to generate a legitimate competition model. For this model to apply to other sports, such as basketball or football, more data becomes necessary. One solution to this problem is the breakthroughs in learning algorithms, or simply by collecting more data.

Although the model can generate two animated characters who can compete with each other, the quality of the input reference also decides the natural degree of action performance.

Although FAIR’s research still has many limitations, it is mostly concerned with coming up with a simulation method that allows humans to interact using AI. Hopefully, this will open up new forms of application in computer games, commercial films, and other sports events.