On Safe Robot Navigation Among Humans as Dynamic Obstacles in Unknown Indoor Environments

In this paper, we rigorously test two conjectures in mobile robot navigation among dynamic obstacles in unknown environments: i) a planner for static obstacles, if executed at a fast update rate (i.e., fast replanning), might be quite effective in dealing with dynamic obstacles, and ii) existing implemented planners have been effective in humans environments (with humans being dynamic obstacles) primarily because humans themselves avoid the robot and if this were not the case, robot will run into collisions with humans much more frequently. The core planning approach used is a Global path planner combined with a local Dynamic Window planner with repeated re-planning (GDW). We compare two planners within this framework: i) all obstacles are treated as static (GDW-S) and ii) predicted trajectories of dynamic obstacles are used to avoid future collisions within a given planning horizon time (GDW-D). The effect of humans avoiding robot (and other humans) is simulated via a simple local potential field based approach. We indicate such environments by a suffix +R (repulsion) for the corresponding planner. Hence there are four categories that we tested: GDW-S, GDW-D, GDW-S+R and GDW-D+R in different environments of varying complexity. The performance metrics used were the percentage of successful runs without collisions and total number of collisions. The results indicate that i) GDW-D planner outperforms GDW-S planner, i.e., conjecture 1 is false, and ii) humans avoiding robots does result in more successful runs, i.e., conjecture ii) is true. Furthermore, we've implemented both GDW-S and GDW-D planners on a real system and report experimental results for single obstacle case.