We are considering the problem of controlling a team of robotic bodyguards protecting a VIP from physical assault in the presence of neutral and/or adversarial bystanders. This task is part of a much larger class of problems involving coordinated robot behavior in the presence of humans. This problem is challenging due to the large number of active entities with different agendas, the need of cooperation between the robots as well as the requirement to take into consideration criteria such as social norms and unobtrusiveness in addition to the main goal of VIP safety. Furthermore, different settings such as street, public space or red carpet require very different behavior from the robot. We describe how a multi-agent reinforcement learning approach can evolve behavior policies for teams of robot bodyguards that compare well with hand-engineered approaches. Furthermore, we show that an algorithm inspired by universal value function approximators can learn policies that exhibit appropriate, distinct behavior in environments with different requirements.
View on arXiv