Modeling Time-Aware Mobile Robot Navigation by Learning Subjective Time Maps (STM)

The basic operation of a mobile robot is navigating to some target, avoiding collisions and possibly minimizing other criteria. A diversity of methods have been developed since the past century, and the research is still active, but there is one aspect that is often neglected: the duration of the steps in which computational devices divide the navigation process. Usually, it is set heuristically to a small, constant value for sampling observations frequently enough to ensure safety; however, each robot and environment has particularities that can make such a fixed timestep sub-optimal under some criteria. This paper explores the possibility of learning the best time for each step in a particular robot-environment interaction, and shows that those subjective, variable timesteps constitute an important aspect of both motion safety and efficiency. The proposed approach has an initial learning stage in a general, minimalistic environment designed for providing the main situations the robot will face; through deep RL, a time-aware reactive navigation policy is found that optimizes collision avoidance, total navigation time, and accurateness in reaching the target, yielding (motion, timestep-duration) commands. Due to the generalization capabilities of DRL, that policy induces Subjective Time Maps (STMs) for time-aware navigation in other scenarios. Thus, STMs become time-aware subjective models of the robot-world interaction that improve that particular robot navigation and also allow it to leverage time effort, e.g., longer timesteps can be used for energy savings or for other operations. Furthermore, the integration of uncertainties in both localization and decision-making is addressed. To achieve this, a time-aware motion policy is first learned while ignoring localization uncertainty, in order to improve learning convergence and reduce computational cost. The policy is then integrated with the resulting STMs through a particle filter (PF), either by using the time-aware motion command provided by the STM on the representative particle or by associating each particle (robot pose hypothesis) and its importance (weight) with it. The goal is to enable a full treatment of uncertainty with reduced complexity. Diverse experiments and comparisons with STMs show the utility of this spatio-temporal representation of knowledge for safe navigation, and the pros and cons of the two methods of uncertainty integration.