Robot Fast Adaptation to Changes in Human Engagement During Simulated Dynamic Social Interaction With Active Exploration in Parameterized Reinforcement Learning

Dynamic uncontrolled human-robot interactions (HRIs) require robots to be able to adapt to changes in the human's behavior and intentions. Among relevant signals, nonverbal cues such as the human's gaze can provide the robot with important information about the human's current engagement in the task, and whether the robot should continue its current behavior or not. However, robot reinforcement learning (RL) abilities to adapt to these nonverbal cues are still underdeveloped. Here, we propose an active exploration algorithm for RL during HRI where the reward function is the weighted sum of the human's current engagement and variations of this engagement. We use a parameterized action space where a meta-learning algorithm is applied to simultaneously tune the exploration in discrete action space (e.g., moving an object) and in the space of continuous characteristics of movement (e.g., velocity, direction, strength, and expressivity). We first show that this algorithm reaches state-of-the-art performance in the nonstationary multiarmed bandit paradigm. We then apply it to a simulated HRI task, and show that it outperforms continuous parameterized RL with either passive or active exploration based on different existing methods. We finally test the performance in a more realistic test of the same HRI task, where a practical approach is followed to estimate human engagement through visual cues of the head pose. The algorithm can detect and adapt to perturbations in human engagement with different durations. Altogether, these results suggest a novel efficient and robust framework for robot learning during dynamic HRI scenarios.