Missteps in Robot Social Navigation

Assessing the quality of robot social navigation is a challenging problem fraught with human obstacles. From preconceived notions to perspective or point of view, evaluations can differ from person to person. Most work in the field of robot navigation is focused on creating algorithms that produce efficient robot trajectories. We posit that the evaluation of trajectories in a social context is essential and distinct to trajectory generation. In this work we recorded a manually driven powered wheelchair through different scenarios and asked expert evaluators to assess the quality of the powered wheelchair's movement. These evaluations were then compared to post-experiment assessments from trajectory generation algorithms and social navigation concepts. Our results show that it is possible to build a simple model to predict expert evaluators' responses. Unfortunately, there is no clear consensus amongst these experts on what quality behaviour is. This suggests that while current navigation algorithms offer strong heuristics for the generation of smooth trajectories in well-defined environments, their efficacy in evaluating social navigation is less obvious. We believe that more emphasis must be put on dynamic and reactive navigation algorithms as any heuristic approach will be limited due to variance in people's behaviours and expectations.