Human Robot Pacing Mismatch

TL;DR

It is argued that a broader cause of suboptimal navigation performance near human is due to the robot's misjudgement for the human's willingness (flexibility) to share space with others, particularly when the robot assumes the human's flexibility holds constant during interaction, a phenomenon of what is called human robot pacing mismatch.

Abstract

A widely accepted explanation for robots planning overcautious or overaggressive trajectories alongside human is that the crowd density exceeds a threshold such that all feasible trajectories are considered unsafe -- the freezing robot problem. However, even with low crowd density, the robot's navigation performance could still drop drastically when in close proximity to human. In this work, we argue that a broader cause of suboptimal navigation performance near human is due to the robot's misjudgement for the human's willingness (flexibility) to share space with others, particularly when the robot assumes the human's flexibility holds constant during interaction, a phenomenon of what we call human robot pacing mismatch. We show that the necessary condition for solving pacing mismatch is to model the evolution of both the robot and the human's flexibility during decision making, a strategy called distribution space modeling. We demonstrate the advantage of distribution space coupling through an anecdotal case study and discuss the future directions of solving human robot pacing mismatch.

Figures (4)

• Figure 1: Example of preference distributions at each time step. The cross is the intent at each time step, while the ellipsoidal circle represents the flexibility.
• Figure 2: The robot's planned trajectory and predicted pedestrian trajectory with three different strategies in the case study. The solid dots are the start position of the robot and the pedestrian. The solid lines with arrows are the predicted pedestrian trajectory and planned robot trajectory. The dashed lines form the envelope of preference distribution. Left: The robot predicts pedestrian trajectory ahead of planning, and does not take into account the influence from itself to the pedestrian. As a result, the robot predicts the pedestrian will leave no space for the robot to pass simultaneously, and robot chooses to dodge the human; Middle: The robot simultaneously predicts pedestrian trajectory and plans its own trajectory. Even though prediction and planning are coupled, the pedestrian preference is measured in open space but is used for close proximity interaction at the door. This static preference assumption leads to an incorrect estimate of pedestrian preference during interaction. As a result, the robot predicts the pedestrian's pacing is to slow down before the door and then accelerate to jump the gap, and robot plans a similarly overaggressive trajectory; Right: The robot predicts the evolution of pedestrian preference at each time step during interaction, while simultaneously plans the optimal preference for itself. Note that the robot predicts the pedestrian adjust the width of preference distribution near the door to leave space for the robot, and robot adjusts its preference as well in response (more details regarding the evolution of preferences during interaction can be found in Fig. \ref{['fig: dso_steps']}). As a result, both the robot and the pedestrian can pass the door at the same time without compromising safety or efficiency.
• Figure 3: Detailed illustration of agent preference evolution during interaction. Note that both the robot and the pedestrian adjust the flexibility near the door with the presence of close proximity interaction. An animated video of this illustration can be found at: https://youtu.be/2GNeBrdHU34.
• Figure 4: One-dimension illustration of coupled prediction and planning in trajectory space and in distribution space. Left: Trajectory space coupling jointly predict and plan optimal position at each time step (solid lines), with fixed preference distributions (opaque curves); Right: Distribution space coupling jointly predict and plan optimal preference distributions at each time step (solid curves), which evolve from the prior preferences (opaque curves).