The company has one research paper and a blog post published, in which it presents a method by which AI agents must learn to control the movements of physical bodies. The process is based on a universal motor control module called “Neural Probabilistic Motor Primitives” (NPMP). It converts motor movements into control signals and is trained offline or via reinforcement learning by recording human (or animal) movement data via trackers.
The team’s research states, “We optimized teams of agents to play simulated football through reinforcement learning. We narrowed the solution space to those plausible movements that we had learned using human movement pattern data.”
The aim of the exercise is not to develop a better footballer or to like it The next web says: “Cristiano Ronaldo has nothing to fear from the robots for now”. Rather, the AI should help figure out how humanoids and other robots can move with the help of agents in a hard-to-control space and how to predict their movements.
At the start of the model training, the AI is just barely able to somehow move its physics-based humanoid avatar around the playing field. But by “rewarding” an agent every time their team scores a goal, the model can get the bots up and running after about 50 hours of training. A few days later, it is then possible to more accurately predict where the ball will fly and how the other agents will react to an actor’s movements.
The research paper states: “The result is a team of coordinated humanoid football players exhibiting complex behavior at different levels, which is quantified through a range of analysis and statistics – including those used in real-life sports analysis. Our work represents a full demonstration of learned integrated multi-level decision making in a multi-agent situation.”
Or as The Next Web sums it up: “This work is pretty awesome. But we’re not sure if it’s a ‘full demonstration’ of something. The model is clearly capable of controlling a physical agent. Judging by the selected GIFs in the blog post, but this work is still in its infancy.”
However, companies like Boston Dynamics, with their machine learning algorithms and pre-programmed choreographies, have shown that such models become increasingly robust over time. The question now is how such adaptive models will develop once they have left the laboratory environment and are used in real robotic applications. (ss)