Adaptive humanoid robot behaviour in a serious game scenario through reinforcement learning

The study presents an adaptive technique that enables a humanoid robot to select appropriate actions to maintain the engagement level of users while they play a serious game for cognitive training. The goal is to design and develop an adaptation strategy for changing the robot's behaviour based on Reinforcement Learning (RL) to encourage the user to remain engaged. Initially, we trained the algorithm in a simulated environment before moving on to a real user experiment. Thus, we first design, develop, and validate the RL strategy in a simulated environment. Subsequently, we integrate the trained policy into the robotic system, allowing it to select the best actions based on the detected user state during real user test. The RL algorithm was designed and implemented to determine an effective adaptation strategy for the robot's actions, encompassing verbal and non-verbal interactions. The proposed solution was first trained in a simulated environment and then tested with 28 users in a mixed-method design study.