Goal-oriented autonomous decision-making for social robots via collaborative interactive inverse reinforcement learning approach

基于协作交互式逆强化学习方法的面向目标的社交机器人自主决策

阅读:1

Abstract

Since the practical constraints of unknown pedestrian goal information, research on inverse reinforcement learning (IRL) applied to social robots has focused on trajectory planning based on current motion direction, other pedestrians, and obstacles. However, social robots typically have clear navigation goals, and the practicality of the aforementioned methods is debatable. Moreover, trajectory prediction at longer distances also poses significant challenges for such trajectory planning methods. Therefore, this paper proposes a goal-oriented autonomous decision-making (GO-ADM) method, which addresses the social compatibility navigation problem of robots through sequential actions at each discrete time steps, eliminating the need to predict pedestrian trajectories. Specifically, this paper defines and collects goal-oriented expert demonstration and propose a collaborative interactive IRL framework for social robots. By exploring both explicit and implicit norms of pedestrian motion in the expert demonstration, goal-oriented reward function was ultimately integrated into GO-ADM to achieve goal-oriented autonomous decision-making for social robots. Additionally, GO-ADM incorporates a penalty function for social safety distance to further ensure the safety of human-robot interaction. Considering the longitudinal dominant navigation task of pedestrians in real-world scenarios, this study constructs a realistic training environment to validate the longitudinal and lateral dominant navigation tasks of the GO-ADM. The results demonstrate that, under both longitudinal and lateral dominant navigation tasks, the robot is capable of making rational autonomous decisions, with the final average deviation from the destination being less than 0.13 m and 0.23 m, respectively, and average deviation rates are less than 0.18% and 1.1%. Furthermore, reduplicate experiments demonstrate that GO-ADM achieves higher success rate with other social navigation and decision-making algorithms. Under severe interference conditions, with maximum noise 0.5 m, the success rate of GO-ADM exceeding 75%, demonstrating strong robustness against unknown interferences.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。