Socially-Aware Robot Navigation by Leveraging Group Detection

Student thesis: Doctoral ThesisDoctor of Philosophy

Abstract

In the past years, the deployment of embodied agents has become more and more common- place in human-populated environments. This trend poses a need for safe robot navigation with higher levels of autonomy, which is one of the key challenges within the area of embodied agent research. Previous research established reliable navigation strategies, however, they do not take into account social aspects sufficiently, which might hamper user acceptance and trust. Studies show that social awareness is an important capability that allows humans to act in an easy-to-understand and predictable manner, which are both crucial competencies in navigating crowded environments. To equip a robot with the same skills, it first needs to recognise social relations in a social setting. A part of this recognition involves the detection of interaction groups which can aid the robot in traversing crowded spaces in a socially-aware manner.

This thesis presents my research on socially-aware robot navigation in crowded indoor spaces. It has two main parts focusing on group detection and socially-aware navigation, respectively. The first part addresses the problem of group detection in crowds based on both third-person and robocentric perspectives. Throughout the thesis, there is a special emphasis on the development of robocentric solutions as this problem has not been thoroughly explored before. One of the main gaps in the literature was the absence of a publicly available benchmark dataset, which was collected from a robot’s perspective, featuring dense social environments and containing annotations for groups. To address this gap, I proposed the Robocentric Indoor Crowd Analysis (RICA) dataset, which can be used for the training and evaluation of methods in the field of socially-aware robot navigation. Using the RICA dataset, I proposed two approaches to the problem of group detection. First, I devised a solution based on Agglomerative Hierarchical Clustering (AHC). This method relies on position and orientation data of detected individuals in a scene and has achieved state-of-the-art performance among unsupervised group detection methods when evaluated on common benchmark datasets as well as RICA. Second, I presented a Graph Neural Network based group detection approach, which, relying on the same input features as my AHC-based method, achieves the best performance yet among supervised solutions. Following the exploration of supervised and unsupervised group detection techniques, in the second part of this thesis, I examine the viability of using group information for improved socially-aware navigation by integrating it into a reinforcement learning-based method, i.e., an advantage actor-critic (A2C) algorithm. In previous works, there was no agreement with regard to how social awareness of navigation methods should be evaluated. To address this gap, I presented a more thorough and unified method for benchmarking the social awareness aspect of existing solutions. This evaluation method implements the most commonly used quantitative metrics and includes a qualitative assessment as well as a Navigation Turing Test.

Taken together, the techniques and findings documented in this thesis present state-of- the-art group detection methods that are applicable to both third-person and robocentric data, and showcase why and how group information can be utilised for socially-aware navigation and the importance it bears.
Date of Award1 Aug 2023
Original languageEnglish
Awarding Institution
  • King's College London
SupervisorOya Celiktutan Dikici (Supervisor) & Matthew Howard (Supervisor)

Cite this

'