Section: New Results

Attention-Based Navigation

Participants : Thierry Fraichard, Remi Paulin, Patrick Reignier.

The domain of service-robots is growing fast and has become the focus of many researchers and industrials alike. Application areas are extremely broad, from logistics to handicap assistance. A large proportion of such robots are expected to share humans' living space and thus must be endowed with navigation capabilities that exceed the standard requirements pertaining to autonomous navigation such as motion safety. In a human populated environment, optimality does not boil down to minimizing resources such as time or distance traveled anymore, the robot motion must abide by social rules and move in a manner which is appropriate.

Most of the approaches proposed so far rely upon the definition of so-called social spaces, i.e. regions in the environment that, for different reasons, the persons consider as psychologically theirs. Such social spaces are primarily characterized using either the position of the person, e.g. “Personal space”  [36] , or the activity he is currently engaged in, e.g. “Interaction Space”  [41] and “Activity Space”  [45] . The most common approach is then to define costmaps on such social spaces: the higher the cost, the less desirable it is for the robot to be at the corresponding position. The costmaps are ultimately used for motion planning and navigation purposes.

While improving upon the standard “non social” navigation methods, this type of approach intrinsically ignores the correlations between interactions as well as the influence of the robot on those interactions. It thus fails to capture several important features of social navigation, such as the distraction and surprise caused to the surrounding individuals. To overcome those limits, we suggest using the psychological concept of attention, which plays a central role when humans navigate around each other. This concept brings a new degree of control over the motion of the robot, namely the invasive and distracting character of the robot motion, which have so far proven hard to tackle with the conventional tools such as social spaces. Beside leading appropriate motion, attention-based navigation enables interaction through motion by predicting the quantity of attention the human will give to the robot.

Building upon a computational model of attention that was earlier proposed in  [47] , we have developed the novel concept of attention field. The attention field is straightforward to define: it is a measure of the amount of attention that a given person would allocate to the robot, should the robot be in a given position/state. It is a mapping from the state space of the robot to $\mathrm{I}\mathrm{R}$. We use this attention field in order to carefully control the degree of distraction caused by the robot to the individuals in its surroundings. By monitoring the variations of attentional resources that it causes, we also control the amount of surprise caused by the robot which must be kept to a minimum since it is a cause of discomfort. Furthermore this approach enables us to tackle more complex situations where more than one person is involved such as the task of delivering a private message to an individual, or else joining a group (an example of interaction through motion). Rather than navigating on a single global costmap, this new approach provides for each path several measures of the distraction and surprise caused by the robot on a given individual. Those quantities are then multi-optimised in order to find a path that satisfies all the given requirements for fulfilling the robot's task as well as minimizing the discomfort for individuals who are not directly involved in an interaction with the robot.

In 2015, we have developed a variant of the well-known differential evolution algorithm which deals with optimizing continuous trajectories under multiple constraints. The performance of our approach is now being compared with trajectories obtained by relying only on social spaces. Besides the traditional qualitative approaches to evaluate the discomfort caused by the robot motion, we work on defining more quantitative measures that would enable us to further validate our approach.