Driving on a busy road, or playing competitive team sports such as football, requires people to make split-second decisions based on an anticipation of what those around them will do next.
Now robots are set to be equipped with the skills needed to perform such real-world tasks, thanks to an EPSRC-funded project that aims to teach them to interact with and anticipate the actions of multiple other agents.
The project, which is led by Dr Varuna De Silva at Loughborough University London and also involves Chelsea Football Club Academy, will use an extensive dataset of player and ball tracking from football and basketball to train machine learning algorithms on what humans would do in such circumstances.
Existing artificial intelligence systems are often trained using a technique known as reinforcement learning, in which they are rewarded for making a desirable choice, and therefore learn the best course of action to take.
However, this training technique is less helpful in a multi-agent situation such as driving or playing football, where it is more difficult to identify an obvious reward for a given action.
So instead the researchers are using a technique known as imitation learning, in which the AI system is trained by observing the actions of experts, in the same way that humans learn new skills by imitating those who have already mastered them.
“What we are doing in this case is looking at years of football data, specifically player and ball tracking data, and using this to see how players behave on the field, and in this way to build a robotic model of a footballer,” said De Silva. “Similarly, if we were to give autonomous vehicles the ability to drive in a situation that they are not used to, we will look at how humans have handled the same situation, to teach the robot a suitable and safe course of action.”
Such AI systems could be used in autonomous vehicles and sports analytics, De Silva said. In sports analytics, the system could be used to measure players’ skills, to identify particular talents. Players could be measured against the benchmark of the computational model, De Silva said.
It could also be used in sports broadcasting, to give commentators a much more realistic idea of what players might be capable of in the tough mental and physical conditions of a competitive match.
Finally, in driverless vehicles, De Silva is investigating the possibility of using camera data from traffic lights and other infrastructure to instruct autonomous vehicles in how the best human drivers tackle difficult scenarios such as complicated junctions.