Data processing: generic control systems or specific application – Specific application – apparatus or process – Robot control
Reexamination Certificate
2002-04-08
2004-02-24
Cuchlinski, Jr., William A. (Department: 3661)
Data processing: generic control systems or specific application
Specific application, apparatus or process
Robot control
C700S246000, C700S250000, C700S253000, C700S254000, C700S258000, C701S001000, C701S024000, C701S025000, C701S036000, C701S213000, C706S045000, C706S903000, C318S587000, C180S167000
Reexamination Certificate
active
06697707
ABSTRACT:
FIELD OF THE INVENTION
The present invention relates to the field of intelligent machines. More specifically, the present invention relates to the field of adaptive autonomous robots.
BACKGROUND OF THE INVENTION
While considered hubris by some, the fundamental urge to understand one's self and the surrounding universe, coupled with the technical challenge and the expected benefits of being able to do so, has motivated research in the direction of creating artificial life. One embodiment of artificial life is the adaptive autonomous robot. An autonomous robot is implicitly goal-directed and capable of operating completely on its own by considering its situation in its environment and deciding what actions to take in order to achieve its goals without human intervention. A robot is adaptive if it is capable of improving its ability to achieve its goals.
The task of building an adaptive autonomous robot is sufficiently complex that research groups have partitioned the problem into several more manageable tasks and have concentrated on solving each task independently of the others. Three tasks or behaviors are considered to be the most difficult in robotics; learning, planning, and world representation.
Initial efforts to implement these behaviors in robots were concentrated on building a complex program that processed environmental information from sensors and generated commands to actuators resulting in behaviors that resembled learning, planning, and abstraction (in order to represent the robot's world, or surroundings) in humans.
Although efforts to build a single, complex control program continue, many of the new and exciting advancements in robotics are based upon the rejection of the notion that complex behavior requires a complex control program. Instead, control is distributed to many interacting autonomous agents. Agents are small programs that act independently of other agents while interacting with the other agents. Complex behavior, such as learning or abstraction, emerge from the interaction of many independent agents rather than being controlled by any one agent.
Mataric and Brooks, “Learning A Distributed Map Representation Based on Navigation Behaviors,” in “Cambrian Intelligence: the early history of the new AI,” The MIT Press, 1999, demonstrated that complex behaviors, such as goal-directed navigation, could emerge from the interaction of simpler behaviors termed “reflexes.” A reflex is an agent that couples an actuator signal to a sensor signal. For example, an avoid reflex may generate a signal to a wheel motor based on a signal from a proximity sensor. If the proximity sensor senses an object within a danger zone of the robot, the reflex generates a signal to stop the wheel motor. Mataric and Brooks showed that starting with only four reflexes, goal-directed navigation could emerge from their interaction. The reflexes, however, were not generated by the robot but required hand-coding by a programmer.
Pfeifer, R. and C. Scheier, “Sensory-motor coordination: the metaphor and beyond,”
Robotics and Autonomous Systems, Special Issue on “Practice and Future of Autonomous Agents
,” vol. 20, No. 2-4, pp. 157-178, 1997 showed that signals from the sensors and actuators tended to cluster for repeated tasks and termed such clustering category formation via Sensory Motor Coordination (“SMC”). Cohen has shown that robots can partition the continuous data stream received from sensors into episodes that can be compared to other episodes and clustered to form an exemplar episode. An exemplar episode is representative of the cluster of several episodes and may be determined by averaging over the episodes comprising each cluster. The exemplar episode is self-generated (by the robot) and replaces the external programmer. As the robot is trained, the robot will identify a set of exemplar episodes that may be used to complete an assigned task. The ability of the robot to identify episodes from a continuous sensor data stream and to create “categories” (exemplar episodes) from the clustered episodes may be considered to be a rudimentary form of robotic learning.
In order to gather a sufficient number of episodes for the identification of categories, the robot must be trained. Training is normally accomplished by a reinforcement learning (“RL”) technique as will be known to those skilled in the art. In one example of RL, the robot is allowed to randomly generate actions while a trainer rewards actions that move the robot toward a desired goal. The rewards reinforce the most recent actions of the robot and over time, episodes corresponding to the rewarded actions will begin to cluster as similar actions are rewarded similarly. The training, however, requires many repetitions for each action comprising the desired task. Therefore, there remains a need for a more efficient method of training a robot.
An autonomous robot must be able to select an action that will lead to or accomplish its desired goal. One known method for robot planning involves a spreading activation network (“SAN”), a set of competency modules (“CM”) that, when linked together, initiate a sequence of commands that the robot may perform to accomplish the desired goal. A competency module includes information characterizing the state of the robot both before (state pre-conditions) and after (state post-conditions) a command to an actuator. Competency modules are linked by matching the state pre-conditions of one CM to the state post-conditions of another CM.
Planning begins by first identifying all terminal CMs, defined as CMs having state post-conditions corresponding to the state of the robot after accomplishment of the assigned goal. The state pre-conditions of each of the terminal CMs are then used to find other CMs having state post-conditions matching the state pre-conditions of the terminal CMs. The process is repeated until the state pre-conditions of a CM corresponds to the present state conditions of the robot.
In one method of searching for the shortest path to a goal, each CM is assigned an activation value determined by CMs in contact (matching endpoints) with the CM. The order of execution is determined by the activation value of each CM where the CM with the largest activation value is executed next.
As the number of CMs increases, the time required to complete the search increases very rapidly and the reaction time of the robot increases until the robot is unable to respond to the dynamic changes in its environment. While such a search may be acceptable for planning before beginning a task, the exponential increase of the search time as more CMs are added (i.e. as the robot learns) renders such a search unsuitable for real-time response to the robot's changing environment.
The back-propagation of CM linking creates an unavoidable delay in the robot's responsiveness because the robot cannot begin to execute the linked CMs until the complete chain of CMs taking the robot from its present state to the goal state are found. This unavoidable delay limits the operating environments of the robots to situations that are usually predicable.
Therefore there remains a need for an efficient method for robotic planning capable of reacting to sudden or dynamic situations in the robot's environment while allowing for the addition of CMs as the robot learns.
In robots, as well as humans, the amount of sensory information received greatly exceeds the processing capability of the robot. In order to function in any environment, a robot must be able to condense the voluminous sensor data stream to a data rate that its processors can handle while retaining information critical to the robot's operation. In one method of condensing the sensor data stream, the robot builds a representation of the robot's environment (the world model) and compares the received sensory information to the representation stored by the robot. The world model allows the robot to orient itself in its environment and allows for rapid characterization of the sensory data to objects in the world model.
The world mod
Cuchlinski Jr. William A.
Marc McDieunel
Pennie & Edmonds LLP
Vanderbilt University
LandOfFree
Architecture for robot intelligence does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Architecture for robot intelligence, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Architecture for robot intelligence will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3330442