The motivation of N1-[SocialSpace] comes from the growing number of robots deployed in human environments and the resulting need for key technologies that enable robots to act socially conform, safely and efficiently under such conditions. The project aims at two such key technologies namely the perception of humans and human-aware navigation. The general approach we take is to acquire models that describe relevant human aspects such as attributes, behaviors, and social relations, and incorporate them as background knowledge into systems for people tracking, socially-aware task and motion planning and human-robot interaction.

Concretely, we pursue two modeling approaches, the first one using Hall's Proxemics theory to find the point at which a novel multi-model hypothesis group tracker works best, and the second one using the social force model, a computational model for crowd simulation, to improve the performance of a multi-hypothesis people tracker through better predictions of human motion. We also take learning approaches, for instance, to acquire a spatio-temporal human activity model based on spatial Poisson processes. This model, called spatial affordance map, turned out to be particularly powerful since it contributed to both areas addressed here. It is able to improve tracking of people through spatially informed data association and better motion prediction of maneuvering targets over occlusions. It also delivers the relevant costs to formulate and solve two novel planning problems for social robots, the maximum encounter probability planning and the minimum interference coverage problem. The latter one was exemplified with a noisy vacuum cleaner that learns to avoid busy places and times in order to minimize annoyance.

Further activities include unsupervised learning of dynamic objects, an issue of high importance in populated environments, where not every object category can be anticipated. Also, we are concerned with the development of principled people detectors for 3D range data and RGB-D data, two sensory modalities that become available only recently. For tracking people in such data, we investigate combinations with new approaches to 3D multi-hypothesis tracking and on-line adaptation. In the area of human-aware navigation, we propose an iterative planner to implement slipstream navigation, a strategy for robots to find and follow persons that share the same motion intention through a crowd. And we have introduced FLIRT features, generalized features for 2D range data inspired by visual interest points, and demonstrated their robustness for navigation tasks such as global localization, loop closing, incremental mapping and SLAM.

Finally, within N1-[SocialSpace], we are interested in the design of robots and the study of novel robot-specific social cues. To this end, we have built the interactive robot platform DARYL, a custom-made, mildly humanized social robot with ten degrees of freedom.