Planning

Planning can be broken down into three questions. Answering the question "Where am I?" is also known as "localization." Localization in PAMM is done with dead reckoning and signposts. Determining the response to "Where am I going?" is "goal selection." PAMM can have predefined goals such as "bring user to cafeteria" or "follow commands from the user via admittance control." The question "How do I get there" is answered by some combination of user-input, path-planning and obstacle avoidance. This last part depends on accurate facility maps. The localization, obstacle avoidance and mapping systems are described below.

A.    Localization

There are two types of localization. Self-localization (also called absolute localization) involves determination of current position without historical knowledge of position. Continuous localization, on the other hand, involves making gradual corrections to a position estimate. For PAMM, several situations exist where it must self-localize. These include re-initializations, accidents or manual relocation of PAMM by facility staff.

Many localization techniques are based on either "range-maps" or "occupancy-grids" generated either from vision or acoustic sensors. While techniques are often effective for continuous localization, they will often fail in self-localization when the environment contains many ambiguous locations. A typical eldercare facility has 2 to 5 floors 2000 sq. meters per floor. Inspections of one facility suggest that enough such ambiguities exist in eldercare facilities to make self-localization via range-maps and occupancy grids difficult.

PAMM currently uses a signpost-based vision-localization system. Signposts, like the one shown in Figure 6, are placed strategically on the ceiling of the facility and an upward-facing camera sends images to the frame-grabber. At any time, at least one signpost is visible, allowing PAMM to determine its position and orientation within an assisted living facility. The signpost has three elements. The first two, "orientation marker" and "centerpiece marker" are self-explanatory. Each signpost also has a unique pattern of identification markers. The presence or absence of individual markers represents a binary number. A design with N placeholders for identification markers allows 2N+1-1 separate signposts.

The signposts must be distributed so that, at any given time, at least one signposts is visible in the camera’s field of view. The region of configurations from which the camera can detect a given signpost is called the "detection window" for that signpost. If a signpost’s WD is 6 sq. meters, then a facility with a total of 10000 sq. meters would need at least 1666 signposts. A design with N=10 would suffice. While installing 1666 signposts might seem like a lot of work, it would still be competitive with other conventional self-localization schemes (e.g. electromagnetic systems).

      Figure 6: Localization Signpost

B.    Obstacle Avoidance

For obstacle avoidance, the admittance controller uses a shared control system similar is used similar to the one used by Aigner and McCarragher. It combines user input and obstacle detection to prevent collisions yet allowing the user to exert control over which obstacle free path is taken.

Control sharing can also be thought of as "control filtering." First, user-input from the force/torque sensor is categorized as "forward", "left", "right", "left forward" or "right forward". The presence of an obstacle within 30" of PAMM is then represented as a three element binary array, S. The three elements correspond to detection of objects by the middle three acoustic transducers. A "S =100", for example, would indicate an obstacle was detected at less than 30 inches forward and to the left of PAMM. The admittance controller filters user input according to Table 3.

Table 3: Shared Admittance Control

User-Input

Allow when S=

Forward

000, 100, 101, 001

Left

All

Right

All

Forward left

000, 001

Forward right

000, 100

The discrete shared control described above was suitable for preventing collisions, but it made PAMM difficult to control in cluttered environments. Currently, for trajectory control, the PAMM stops when it detects an imminent collision and waits for its path to clear. A method is being implemented for replanning paths when an obstacle is present.

C.    Mapping

PAMM will have a fairly accurate map of the facility including walls, doors and immovable furniture. A Dempster-Shafer technique described in [ii] is being evaluated for fusing acoustic readings from new objects into evidential grid maps. With this technique, evidence of occupancy and emptiness are accumulated in separate grid maps. These evidential maps can easily be adapted for use in obstacle avoidance algorithms.

One major function of the planner is to select goals. The PAMM might choose to guide the user to a chair. In order to accomplish this goal, it will be necessary for PAMM to know where the chairs are. One solution to this problem might be to mark the chairs so that a vision system can easily distinguish them from other objects. This solution, however, would increase the facility’s costs necessary to accommodate PAMM’s.

Some of our recent experiments suggested that the acoustic sensors might be able to recognize certain objects. The concept, shown in Figure 7, involves "listening" to the echoes from a "target". A chair, for example, would generate a different echo than an object such as a wall, a pillar, or a slim person. Initial results are promising, and more studies are being done to determine the limits this approach to object recognition.

Figure 7: Acoustic object recognition.