The ultimate performance metric for this or any commercially viable automated system is its total benefit over the total cost of ownership. The robot must perform comparable to a human operating in the same task domain. The total cost of capital and maintenance over the robot's serviceable lifetime must be less than the total salary of a human worker over the same time period. However, from the perspective of the robot, and not the perspective of a business manager, performance metrics may include the amount of distress calls (requests for human assistance), the total area cleaned during its shift, and the amount of refuse remaining after the robot is satisfied that a room is clean. In the above cases, the robot should minimize the number of distress calls, minimize the amount of residual refuse, and maximize the total area cleaned.

To keep the cost of initial investment low, the robot should be able to operate within the office without expensive modifications to the building. This means no installations of subfloor guidance or complicated wireless beacons for navigation. To minimize the total cost of ownership, the robot should exhibit a high degree of autonomy. A robot that must operate without careful human supervision in the confines of a clustered office building must exhibit a high degree of maneuverability. It must be able to move accurately, avoiding delicate obstacles that could be toppled by a collision. Furthermore, the tasks for which the robot is designed require dexterous manipulators and sufficient storage for refuse.

From these requirements, I propose a small, wheeled robot of approximately 2 feet in width, 3 feet in length, and roughly 1.5 feet in height. The wheels should be built of an appropriate dimension and of a soft rubber that has good traction properties on both high traffic carpet and buffed tile floors.

The robot should carry a removable bin for easy emptying of collected refuse. Interchangeable bins could be magnetically locked onto the robot. To replace a bin, the robot would back into a docking port and begin communicating with the host computer. The robot would release its magnetic lock, and the host controller would engage the magnetic lock on the docking port. The robot then need only move out of the docking port to free itself of a full bin. To acquire a new bin, it would back into another docking port and use the magnetic locks to secure the bin to the carriage.

The proposed robot can be quite heavy and must operate for extended periods of time (an estimated 8 to 16 hours). Current energy storage techniques are not likely to be cost effective. We can couple the primary power source of the robot to the waste bin. When the robot exchanges bins, the empty bin contains a freshly charged energy cell. A secondary power cell, sufficient for a few minutes of operation for maneuvering between bins, is embedded in the robot and can be charged through the primary power source. Human custodians can replace either the bin or the power cell by decoupling and replacing the faulty component.

While in the docking port, the host controller and the robot can communicate about tasks that remain in the queue, the status of other robots that may be operating, and any other special circumstances which may have arisen since the last time the two were in communication. This communication model allows some centralized control without the expense of installing wireless networking hardware throughout the facility in which in the robot operates.

If the robot is to operate on many floors throughout the office building, some alterations to the facility may be needed. I recommend enhancing the elevators with an infrared communications port similar to those used by television sets. To change floors, the robot moves to an elevator and sends a signal to the infrared port near the elevator. When the elevator has arrived it moves into the elevator and monitors an infrared port in the elevator. When the elevator has arrived at the desired floor, the elevator sends an infrared signal to the robot. The robot then moves out of the elevator when the doors are opened.

Constructing the robot in a modular fashion by using components, each with some degree of autonomy, allows the manufacturer to construct new robots using in less time and cost than a completely new design. Furthermore, microprocessors have become inexpensive with respect to their computational power. Using a task planner in each component has little additional cost and distributes the computational load, making the robot more responsive to its environment. Thus, I propose constructing the robot of three main components. Each component has sensors, affecters, and task planning capabilities. The carriage component is responsible for moving the robot. The manipulator component performs fine motor tasks for collecting refuse. The entire system is guided by a third component, the central planning component.

The carriage consists of the flatbed frame, the drive system, and a task planner. Each of the four wheels has an integrated independent drive system that is mounted on a swivel platform. The wheels can turn at right angles to the center line to allow the robot lateral movement for maneuvering in constrained spaces. Each wheel contains slippage sensors and a proprioceptor for the swivel platform. This information is used by the carriage task planner to adjust the position and torque of the wheels under adverse conditions. The fore and aft of the carriage are equipped with infrared emmitter/detector pairs. These infrared sensors scan the areas immediately to the front and rear of the carriage for continuity in the travel surface. In particular, these sensors inform the carriage task planner if there is a stair drop off in its immediate path. The carriage is also surrounded by a simple bumper. The bumper is partitioned into eight zones. Should something bump into the robot, or should the robot accidentally bump into something due to its actions, the bumper would provide information to the task planner to move the robot away from the stimulus.

The waste bin that fits onto the robot carriage is also equipped with sensors to detect the current load in the bin. These sensors probably consist of linear transducers to detect weight, and optical sensors to detect the fill level in the bin. These sensors and other data about the current state of the primary power cell are fed into the carriage via conducting pads on the surface of the carriage.

The manipulator component consists of a pair of anthropoid arms and a task planner. Each joint in the arm is coupled with a proprioceptor to provide feedback to the task planner. For the given task, the anthropoid arms need not be equipped with humanoid hands. Two fingers and a thumb should be sufficient to grasp waste bins and refuse. The tip of each digit is a soft rubber gripping surface and a linear transducer. The transducer provides the local task planner with feedback about the force being exerted in gripping actions. The tip of each digit also contains a small metallic contact that can detect the presence of metallic objects in its grip. Such information can be used by the robot to sort paper and aluminum refuse into separate bins for recycling.

The central planning component acts as the consciousness of the robot. The central planning unit controls a swivel mounted pair of CCD cameras that feed a stereoscopic image processor. Edge detection and ranging are done in hardware and output is written via DMA into the main memory of the central task planner. The central task planner is responsible for sweeping and scanning the visual field for waste bins and discarded cups while it is performing its routine cleaning operations. Goals of the central planner can be delegated to the carriage or the manipulator. For instance, in planning a travel route, the central planner generates a plan and passes it to the carriage. The carriage is then responsible for moving the robot along the designated path, autonomously making minor corrections to the path while in route. Similarly, if an object is located near enough to be grasped, the central planner passes a generated plan to the manipulator component which is then responsible for executing the plan and making minor changes as needed. Both the manipulator and carriage may request data extracted from the visual field. All such requests are submitted to the central planner which must then position the CCD cameras within the real time constraints provided with the request.

The central planner may also use visual cues to orient itself with respect to a provided floor plan model of the building in which it operates. These visual cues can be placed in aesthetically appealing holographic wall art. At some depth of the hologram information is encoded that the robot can then match against its location cue database.

In operation, the robot is given a floor plan of the building and a patrol route. The patrol route includes estimated locations of waste bins. The robot leaves its bin at a specified time, probably after the normal staff hours of the building and begins its patrol. The central planning component carries a list of the goals that encode the patrol route. The central planning unit creates a rough travel plan by plotting an approximated Hamiltonian circuit of the objects on its patrol route. The robot then begins its patrol route by creating a plan to the first object in the circuit (such as a cubicle). The robot finds a reference point in its current visual field and plots a rough travel plan from its current location to the reference point. This plan operates in three-dimensional space and potential routes that do not provide sufficient clearance are pruned. After generating a rough travel plan, the central planner passes the plan to the carriage component.

The carriage component then refines the submitted plan from move from here to there via these references to a sequence of motor drives and positioning of the wheels. The carriage component makes minor modifications to this plan while in route, using its infrared look ahead sensors and its bump sensors to update its model of the world. Once the goal location has been achieved a message is sent to the central planner to inform the central planner of its arrival.

While in transit, the central planning component uses the CCD cameras to look for refuse. Should refuse be found, the central planner interrupts the carriage and submits a new rough plan to move the robot into proximity of the refuse. The carriage then executes this new task.

Once near to the refuse, the central planner issues a rough plan to the manipulator component. The manipulator component then refines the plan as it is executed, using the arm proprioceptor and finger transducers to adjust arm positioning and gripping force. Once successfully gripped, the manipulator unit then generates a plan to place the refuse over the waste bin and release the refuse. The conductance sensors can be used to place aluminum cans in a separate area of the waste bin.

After an interruption, the central planner then creates a new plan to return to its original patrol route. This process continues until the robot has completed its patrol, the waste bin is full, or the robot has detected a critical shortage of primary power. In the first case, the robot returns to its docking port and powers down. In the latter two cases, the robot returns to a docking port and picks up a new bin.