Robots in multi-user environments require adaptation to produce personalized interactions. In these scenarios, the user’s feedback leads the robots to learn from experiences and use this knowledge to generate adapted activities to the user’s preferences. However, preferences are user-specific and may suffer variations, so learning is required to personalize the robot’s actions to each user. Robots can obtain feedback in Human–Robot Interaction by asking users their opinion about the activity (explicit feedback) or estimating it from the interaction (implicit feedback). This paper presents a Reinforcement Learning framework for social robots to personalize activity selection using the preferences and feedback obtained from the users. This paper also studies the role of user feedback in learning, and it asks whether combining explicit and implicit user feedback produces better robot adaptive behavior than considering them separately. We evaluated the system with 24 participants in a long-term experiment where they were divided into three conditions: (i) adapting the activity selection using the explicit feedback that was obtained from asking the user how much they liked the activities; (ii) using the implicit feedback obtained from interaction metrics of each activity generated from the user’s actions; and (iii) combining explicit and implicit feedback. As we hypothesized, the results show that combining both feedback produces better adaptive values when correlating initial and final activity scores, overcoming the use of individual explicit and implicit feedback. We also found that the kind of user feedback does not affect the user’s engagement or the number of activities carried out during the experiment.

多用户环境中的机器人需要适应才能产生个性化交互。在这些场景中，用户的反馈引导机器人从经验中学习，并使用这些知识来生成适合用户偏好的活动。然而，偏好是特定于用户的，并且可能会发生变化，因此需要学习来为每个用户个性化机器人的动作。机器人可以通过询问用户对活动的看法（显式反馈）或从交互中估计意见（隐式反馈）来获得人机交互中的反馈。本文提出了社交机器人的强化学习框架，利用从用户获得的偏好和反馈来个性化活动选择。本文还研究了用户反馈在学习中的作用，并询问将显式和隐式用户反馈相结合是否会比单独考虑它们产生更好的机器人自适应行为。我们在一项长期实验中对 24 名参与者进行了评估，他们被分为三种情况：(i) 使用通过询问用户对活动的喜爱程度获得的明确反馈来调整活动选择； (ii) 使用从用户行为生成的每个活动的交互度量中获得的隐式反馈； (iii) 结合显性和隐性反馈。正如我们假设的那样，结果表明，在关联初始和最终活动分数时，结合两种反馈会产生更好的适应性值，从而克服了个人显式和隐式反馈的使用。我们还发现，用户反馈的类型不会影响用户的参与度或实验期间进行的活动数量。