|Introduction||A Potential Field Solution||Components of the Approach||Experiments|
|Papers||Funding Details||Contact Details|
The primary goal of this research is to develop mechanisms for allowing a team of robots to carry out stealthy tasks in the presence of a single or multiple observers in outdoor environments. These tasks may be as simple as navigating to a goal or more complex such as surrounding a location/observer to form a sensor network. For these type of tasks, the robots may be carrying out operations within sensor range of the observers. Our initial approach is directed towards evaluating the reduction in exposure between team members as a result of sharing task-related information. The task chosen to demonstrate this approach is a low-visibility traverse to a goal location in the presence of a single observer with a known location. This is carried out with Pioneer AT robots in an outdoor environment that contains barricades for the robots to hide behind during their traverse. However, the robots have no a priori environment knowledge and must develop a map of the environment whilst keeping a low profile. The robots carry out their traverses sequentially with each benefiting from the experience of the previous to improve its performance.
The approach is based on using potential fields to determine good waypoints for a robot carrying out its traverse. Potential fields provide a medium for combining features of the task and environment. The result is an virtual feature space which can be analysed to determine task sub-goals. For our application, the features encoded in the potential fields are based on an occupancy grid each robot develops or maintains and the locations of the robot, goal and observer. Combining these fields produces an abstracted environment representation with embedded task knowledge from which a waypoint is extracted. From the manner in which the potential fields are modelled, the waypoint is typically in either a position obscured by the observer, or in an open, detectable space towards the goal location. The waypoint is the next subgoal for the robot until it changes on subsequent evaluations.
Once a robot has completed its traverse, it has its version of the environment encapsulated as an occupancy grid and a record of interesting waypoints along its path. The robot's path is analysed en route to generate these waypoints of interest which effectively summarise the path. When a robot is about to commence its traverse, it requests the occupancy grid and these waypoints from the preceeding robot. This information is used as a baseline for the robot's traverse and assists in waypoint calculations. As new environment information becomes available during the traverse, the robot overwrites the existing occupancy grid since each robot will present a different representation of the world. This allows more ego-centric decisions to be made but still use the older information for path planning. As such, it is anticipated that the relative locatlisation and sensor view between robots is not grossly different. Sharing information between robots results in more efficient and lower visibility paths.
The types of potential fields used in the approach are shown in Figure 1. The red dots repsent the source of creation. Each image represents a 3-d view of a potential field with peaks being lighter. High areas are generally areas the robots will not traverse to and are considered 'repulsive'. Figure 1b shows an attractive potential field with the most attractive point being at the source. Figure 1c shows how shadow potential fields are formed by generating a valley behind an object from the observer's position.
Figure 1a) Repulsive field. b) Attraction field. c) Shadow field on a repulsive field
These fields are generated from the information in the occupancy grid and about the task. Occupancy grid information consists mainly of the objects that have been found in the environment. Task information consists of the locations of the observer, goal and robot.
Combinations of these fields are used to determine the currently 'best' waypoint for the robot to travel to. Each field is weighted by an empirically adjusted parameter to determine the effect that abstract feature has on the waypoint selection process. For example, a heavy weighting the goal location potential field will cause the robot to converge to the goal quicker, and ignore the observer more than if it was weighted the same as the other fields.
Figure 2 shows the potential field view of the first robot during its traverse in an environment measuring 55m x 55m. The goal is the yellow dot on the right hand side, the observer's location is the red dot at the bottom of each image, the robot's start location is on the left side as indicated by the beginning of the track with the robot being at the end of the track. The robot's next waypoint is shown as the disassociated green dot in 2a and the brown dot at the end of the red waypoint track in 2b. The robot's path indicates places where teh robot assumed it was being detected as lighter colours than where it was undetected.
Figure 2b shows a similar traverse as 2a with the history of the waypoints shown in red. Each waypoint was generated with the limited environment object information available at the time, hence some appear inside of objects that did not exist at the time of calculation. Click on the figures for a larger, clearer image.
Figure 2a) Potential field with superimposed robot path. b). Similar image with the waypoint track.
Experiments have been conducted in simulation and the real environment. The simulation experiments were carried out using the Stage simlator and Player interface. Each environment was 35m x 35m with several barricades for the robots to hide behind. The configuration of the first environment is shown in Figure 3a. The robots start on the lower left side and traverse to the goal (yellow) in the presence of the observer (red). They initially have no knowledge of the environment objects and consequently, the first robot's traverse is highly reactive to the objects in the immediate vicinity. An example of its travserse is shown in blue in Figure 3b for simulation and in Figure 4a for the real environment. The subsequent robot traverses benefit from the information sent by their predecessors by producing better paths as examplified by the second (green track in Figure 3b, Figure 4b) and third (blue-green track in Figure 3b, Figure 4c) travserses. In each of the tracks, the lighter coloured areas indicate where the robot assumed it was being detected.
Figure 3a) Environment configuration for experiment 1. b) An example simulation trial superimposed over the occupancy grid of robot 3. The occupancy grid shows objects in white, explored space in black and unexplored in grey.
Figure 4. An example of the three real robot traverses for experiment 1.
Fast and furious movies of the real robot traverses for experiment 1:
The second experiment demonstrates the robots' performance in a dynamic environment. The configuration of this environment is shown in Figure 5a. The barricade around the observer (red) is added after the first robot's traverse and allows a zero-visibility path to the goal (yellow). Without the barricade, the first robot's traverse becomes a direct path to the goal in almost full view of the observer as shown in blue in Figure 5b and in the left image in Figure 6. The barricade is placed around the observer prior to the second robot's traverse. Initially, it will attempt to follow the first robot's path until it discovers the new barricade. Since this presents a lower-observable path, it changes direction and travels around the barricade and then to the goal as shown by the green path in Figure 5b and the centre image in Figure 6. The third robot's traverse capitalises on the second robot's path to follow one that is more efficient (the yellow path in Figure 5b and the right image in Figure 6).
Figure 5a) Environment configuration for experiment 2. b) An example simulation trial superimposed on the occupancy grid of robot 3.
Figure 6. The occupancy grid view of the real robot traverses for experiment 2. Note that the large horizontal object to the lower left of the goal in the left figure is an artifact added by an incorrect sensor reading whilst the robot was moving behind the goal.Consequently, it did not affect the path.
Movies of the real robot traverses for experiment 2:
And if you want truly stealthy operations:
Ashley Tews, Gaurav S. Sukhatme, and Maja J. Mataric'. A Multi-robot Approach to Stealthy Navigation in the Presence of an Observer. To appear in IEEE International Conference on Robotics and Automation (ICRA), 2004. pdf
This project is funded by DARPA
Grant Description: MARS2020 via UPenn
Contract Number: 5-39509-A