Towards Assessing the Human Trajectory Planning Horizon

Abstract

Mobile robots are envisioned to cooperate closely with humans and to integrate seamlessly into a shared environment. For locomotion, these environments resemble traversable areas which are shared between multiple agents like humans and robots. The seamless integration of mobile robots into these environments requires accurate predictions of human locomotion. This work considers optimal control and model predictive control approaches for accurate trajectory prediction and proposes to integrate aspects of human behavior to improve their performance. Recently developed models are not able to reproduce accurately trajectories that result from sudden avoidance maneuvers. Particularly, the human locomotion behavior when handling disturbances from other agents poses a problem. The goal of this work is to investigate whether humans alter their trajectory planning horizon, in order to resolve abruptly emerging collision situations. By modeling humans as model predictive controllers, the influence of the planning horizon is investigated in simulations. Based on these results, an experiment is designed to identify, whether humans initiate a change in their locomotion planning behavior while moving in a complex environment. The results support the hypothesis, that humans employ a shorter planning horizon to avoid collisions that are triggered by unexpected disturbances. Observations presented in this work are expected to further improve the generalizability and accuracy of prediction methods based on dynamic models.

Multi-agent path topology in support of socially competent navigation planning

“…Plans have a short horizon but are made with a global reasoning (over joint strategies). The short planning horizon has been shown to be in compliance with human locomotion according to Carton et al (2016a), who presented evidence that humans employ a shorter planning horizon as they navigate complex environments, to avoid collisions that could emerge from unexpected disturbances. The global planning horizon ensures that the motion of the robot will be consistent throughout the whole sequence of consecutive planning cycles.…”

“…This is enabled through a sophisticated mechanism of perception and action, enabled through information exchange mostly via path shape, body posture, and gaze (Goffman, 1966) that has been widely studied from a number of different fields. For example, Carton et al (2016a) studied the trajectory planning horizon of humans in locomotion tasks towards informing the design of models for the prediction of human walking behaviors. In the field of psychology, Warren (2006) proposed a model that may describe organization in human behavior in a number of tasks by treating an agent and its environment as a pair of coupled interacting dynamical system.…”

Risk-aware motion planning for automated vehicle among human-driven cars

“…is defined in (20), C th is the cost threshold below which human drivers are not sensitive to, and p th is the corresponding probability threshold. Note that (23) only considers the time horizon of one step ahead, which corresponds to the findings in [16] that the look-ahead horizon of human drivers are relatively short in uncertain decision-making settings.…”

Risk-sensitive inverse reinforcement learning via semi- and non-parametric methods

“…As we model each participant as planning over a receding horizon with two decision stages, it is reasonable to assume that the plan for the second stage is not as fine-grained as over the first one. In addition to reducing the computational burden, this concept of mixing coarse/fine planning is a feature also described in Carton et al (2016) to model human locomotion. We observe that using 15 control trajectories for the first stage and 5 trajectories for the second stage was suffcient to generate a diverse expert action set in terms of accelerations (Figure 11a and (c)) and steering rates (Figure 11b and (d)); the span of all x ∕ y traces resulting from the combination of these first and second stage control trajectories is shown in Figure 12.…”