|Introduction||Activity Data||Spatial Features||Occupancy Grids||Publications||Support||Contact|
In this project, we study the activities of groups of people and robots and analyze the spatio-temporal patterns exhibited in these kinds of interactions. Our goal is to derive models that characterise these activities and apply these models in generating group-level primitives such as social primitives. Our approach is partly based on ideas from social psychology theories such as proxemics (the study of space). Our belief is that by applying these well-established theories, we may be able to generate social behaviors that more closely resemble human's.
In order to perform our analysis, we need to gather a significant amount of activity data. Our activity data consist of sequences of (x, y) positions of robots or people over long periods of time under different experimental settings. These activity data are collected using our planar laser-based people/robot tracking system. We also collect simulated robot activity data using the robot simulator Stage.
We apply ideas from social psychology theories such as proxemics (the study of space) in deriving a set of spatial features to help us understand the relationship between behaviors and their spatial characteristics. We then analyze the characteristics of these features under different experimental and feature parameters.
Overview of Spatial Features
The spatial features we use in our analysis are:
We validated the above spatial features using robot position data collected for multi-robot flocking, random-walk and boundary-tracing behaviors under different experimental parameters (using our robot server and simulator Player/Stage). For flocking, we observed formation of mostly one cluster from the entire group, each robot being very close to one another. For random walk, the robots distributed evenly throughout the given space, leading to maximum isolation. Boundary-tracing robots, on the other hand, cluttered along the edges of the environment, causing fewer cluster formations than flocking robots, yet less isolated than random-walking robots.
Here are some short video clips showing the spatial variations of 8 robots random-walking, boundary-tracing and flocking (robots are denoted by circles, clusters are outlined by dashed lines) (using 1.1m as clustering threshold):
For small-group human activities, we collected position data for unscripted activities over a long duration using our laser-based people-tracking system. Here is a video clip showing spatial variations of people in our lab (people are denoted by circles, clusters are outlined by dashed lines) (using 1.5m as cluster threshold):
Temporal occupancy grids extend the concept of occupancy grids in time dimension. In addition to the probability of occupancy, the speed and frequency information are encoded into each grid, for instance people tend to move faster in regions near to a door and slower in waiting areas in front of the elevators. We can use the temporal occupancy grids to infer the "pace" of activity at different regions of an environment by post processing at different temporal resolutions.
Here is an example of the temporal occupancy grids of our lab.
This work is supported by DARPA Grant DABT63-99-1-0015 under the Mobile Autonomous Robot Software (MARS) program, and in part by the ONR MURI Grant (with UC Berkeley, Stanford, and Caltech).