Interaction Lab

Introduction and Overview	Approach	Environmental Complexity Considerations	Results
Movies	Related Papers	Support	Contact Details

Top

Introduction and Overview

While typical human-robot interaction (HRI) focuses on one-on-one (or at least few-to-few) tasks, there are a host of additional applications that consider multiple robots interacting with multiple people. One example is the evacuation assistance domain, wherein robots are intended to interact with "flows" of evacuating pedestrians. Having robots interact with groups of people, who are themselves interacting, is a unique and challenging problem because robots need to be able to manage and even exploit the crowd's own intrinsic dynamics (in addition to the uncertainty, physical dynamics and robot-to-robot effects).

A community of researchers focuses on better understanding the complex relationship between people and the environment. Techniques including models and simulations are used to calculated the effect of environmental changes on egress time. We have begun to use these pedestrian and evacuation dynamics methods for robotics, both as on-line methods within the algorithms, and for experimental evaluation.

This work is an example of robotic disaster response work, but differs from most existing research focusing on the Urban-Search and Rescue problem.

Other work in the Interaction Lab lab relates to this project:

The data driven approaches used to understand spatial interaction behavior central to our Activity Modeling work. While both predictive and generative models have been used here, predictive models alone may be sufficient. Typical interactions studied with automated activity modeling techniques are of a much smaller scale (spatially), and thus the robot controllers synthesized in those cases would be that type of activity -- probably what would be called "social" rather than "mass."
The mechanisms used to coordinate the robots in order to ensure that deployment occurs in a minimally interfering manner, etc, is based on prior work in Multi-Robot Coordination and Learning. The exploitation of intrinsic dynamics in a many body processes need not only be applied to people. (Although it should be noted that this is arguably the only way that control could be accomplished for people.) If synthesis can result in robots with well understood, or even predictable dynamics, then the general methods that this work points toward.
The work also shares a philosophical foundations with notions like Behavior-based control and Motion Primitives. Fundamentally they seek parsimonious factorizations of behavior that capture the regularity afforded by the real-world constraints. Methods like dimensionality reduction are techniques that permit this structure to be identified, qualitatively understood, and used. We claim that interacting with people, and other entities with dynamics of their own, requires acquiescency with the importance of those dynamics. Addressing questions regarding sets of these structures, means that we are venturing into higher-level compositions, and ideas like basis behaviors become pertinent.

Top

Approach

We have designed, implemented and evaluated a distributed algorithm for the self-deployment of a team of mobile robots intended to carry audio beacons. These directional audio beacons have been shown to assist people in the route finding phase of evacuation, making navigation toward suitable exits efficient.

In order to decide exactly where beacons should be deployed, we used two pieces of information: 1) the current standard practice of deploying beacons at (or near) exits, stairwells, and other strategic locations; 2) an environmental complexity measure shown to correlate with the densities of pedestrians found within a built environment. These two factors are used in a utility calculation, which is used in the multi-robot task allocation phase of the deployment algorithm.

		Directional audio beacons are used in two particular ways, depending upon whether they are positioned at a suitable emergency exit (as in Figure 1a), or near an stairwell on a useful exit route (Figure 1b). In the first case, beacons simply emit a signal that is useful for directional localization. Their sole purpose is as a target. The case of stairways is slightly different because the beacons do not signify a final destination, but part of a possible path toward safety. Thus, in addition to the directional sound, stairway node may also play either a rising or sinking harmonic tone depending on whether evacuees should ascend or descend the stairs respectively.
Figure 1a: Suitable Fire Exit.	Figure 1b: Connecting Stairway.

Tests of beacon effectiveness have shown that subjects correctly interpreted the audio cues, even without any briefing as to the meaning of sound or the signals used at stairwells. Throughout the experiments not a single subject either took a wrong turn or got lost, while the sound overwhelmingly reduced egress time.

Top

Environmental Complexity Considerations

In order to evaluate the complexity of a particular segment of the environment, and more specifically, to compare two different segments, we calculate a Space Syntax measure. This requires the extraction of a skeleton of the environment; typically this is done either with convex polygons, or an axial map. We employ topological information, constructing a graph, and calculating these measures on this graph.

Figure 2 shows a 2.5 dimensional representation of a multi-story building used in simulation experiments. Emergency exits, and topological information is shown overlaid. Robotic beacons were deployed, first five from the the ground floor, and then later one from second floor. Although the robots coordinate in order to assign themselves to locations spread far apart.

Figure 2: The two-story building used in simulation experiments.

Top

Results

Figure 3: Decreasing mean and variance in egress distance (and hence time) with the addition of beacons.

Figure 4: The effect of variation of the bias parameter.

We have evaluated our implementation along two dimensions: 1) robustness with respect to failure and 2) effectiveness in terms of impact on the evacuees. The first case can be investigated through the use of carefully thought-out test cases, the introduction of simulated failures, and validation on physical hardware. These steps were all performed.

The second dimension presents new challenges. Complete evaluation of the effectiveness requires human participants and recreating a simulated emergency. In order to bypass these obvious difficulties, we implemented an evacuation simulation (based on an existing "maximum flow" macroscopic-model). We added a bias parameter that captures the strength of the bias effect on evacuees.

Figure 3 shows that the evacuation simulation suggests that adding a single beacon is highly effective, but thereafter additional beacons are less beneficial. Both means and variances are decreased, which is significant because a decrease in only one has limited value. Figure 4 shows the effect of a range of values for the bias parameter, and that the trend is unaffected.

Top

Movies

Induced failure in PHE [AVI, 8.4MB]

Top