DECISION MAKING MECHANISM, METHOD, MODULE, AND ROBOT CONFIGURED TO DECIDE ON AT LEAST ONE PROSPECTIVE ACTION OF THE ROBOT

• US 20100262286A1
• Filed: 04/02/2010
• Published: 10/14/2010
• Est. Priority Date: 04/03/2009
• Status: Active Grant

• Current Assignee: Siemens AG
• Original Assignee: Siemens AG
• Inventors: Zöllner, Raoul Daniel, Eidenberger, Robert, Grundmann, Thilo
• Granted Patent: US 8,583,284 B2
• Time in Patent Office: Days
• US Class Current: 700/246
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks

Abstract

A decision mechanism is configured to decide on at least one prospective action of a robot from set of actions by: computing a prior probabilistic representation of a prior environment state; updating the prior probabilistic representation with targets of a new observation on reducing at least one uncertainty in a posterior probabilistic representation of a posterior environment state to be reached after an appliance of the at least one prospective action, the posterior probabilistic representation resulting from the updating; determining an information gain between the prior probabilistic representation and the posterior probabilistic representation by use of at least one information theoretic measure; evaluating the at least one prospective action by adding costs of executing the at least one prospective action to the information gain. Furthermore, an improved action planning for robots is provided and can be implemented in various robots investigating scenes for their actions.

Claims

1. A decision making mechanism, said decision making mechanism being operable to decide on at least one prospective action of a robot from a set of actions of said robot by:

computing a prior probabilistic representation of a prior environment state;

updating of said prior probabilistic representation with targets of a new observation on reducing at least one uncertainty in a posterior probabilistic representation of a posterior environment state to be reached after an appliance of said at least one prospective action, wherein said posterior probabilistic representation is a probabilistic representation resulting from said updating;

determining an information gain between said prior probabilistic representation and said posterior probabilistic representation by use of at least one information theoretic measure; and

evaluating said at least one prospective action by adding costs of executing said at least one prospective action to said information gain.

2. The decision making mechanism according to claim 1, wherein said at least one prospective action is an action to be executed by said robot.

3. The decision making mechanism according to claim 1, wherein said at least one prospective action is a robot actuation command demanding a change of position of a sensor of said robot.

4. The decision making mechanism according to claim 1, wherein said decision making mechanism is configured to evaluate a sequence of prospective actions of said robot.

5. The decision making mechanism according to claim 1, wherein said at least one uncertainty comprises at least one of:

a state transition uncertainty arising when reaching said posterior environment state from the prior environment state by said appliance of said at least one prospective action; and

a measurement noise disturbing a measurement performed at said posterior environment state.

6. The decision making mechanism according to claim 1, wherein said decision making mechanism is configured to perform said updating in a Bayes filter.

7. The decision making mechanism according to claim 1, wherein said at least one information theoretic measure is entropy on a probabilistic representation.

8. The decision making mechanism according to claim 1, wherein said determining comprises determining of a difference between a prior entropy of said prior probabilistic representation and a expected entropy of said posterior probabilistic representation.

9. The decision making mechanism according to claim 1, wherein said updating comprises an estimating of the new observation, which will be obtained after executing of said at least one prospective action.

10. The decision making mechanism according to claim 9, wherein said updating comprises predicting of said posterior probabilistic representation of said posterior environment state for said at least one prospective action by use of said estimating.

11. The decision making mechanism according to claim 1, wherein of said prior probabilistic representation and said posterior probabilistic representation are multivariate Gaussian mixture distributions.

12. The decision making mechanism according to claim 1, wherein said at least one uncertainty is a state uncertainty determined by approximating differential entropy of said posterior probabilistic representation by an upper-bound estimate.

13. The decision making mechanism according to claim 1, wherein said decision making mechanism is configured to decide on said at least one prospective action if, after executing of said at least one prospective action, said information gain is maximized and said costs of executing said at least one prospective action and said uncertainty are reduced.

14. The decision making mechanism according to claim 1, wherein the decision making mechanism is configured to terminate said deciding if a desired quality criteria is reached.

15. The decision making mechanism according to claim 13, wherein said desired quality criteria is provided in form of at least one distribution entropy.

16. The decision making mechanism according to claim 1, wherein said prior probabilistic representation and said posterior probabilistic representation are computed by use of a model specified by use of a POMDP concept.

17. The decision making mechanism according to claim 1, wherein said prior environment state and said posterior environment state describe a state of at least one object, wherein said state of at least one object contains a discrete class representation of a class of said at least one object and a continuous m-dimensional pose of said at least one object.

18. The decision making mechanism according to claim 1, wherein said prior environment state and said posterior environment state refer to a state of a scene comprising at least one object located in said scene and determine visibility of features said at least one object in said state.

19. The decision making mechanism according to claim 1, wherein said new observation is expressed as a detection of a set of features of at least one object of a scene.

20. The decision making mechanism according to claim 1, wherein said decision making mechanism is implemented in said robot.

21. The decision making mechanism according to claim 1, wherein said decision making mechanism is implemented in an active perception module of said robot, said active perception module being configured to autonomously investigate scenes.

23. An active perception module comprising a decision making mechanism according to claim 1, wherein said active perception module is configured to autonomously investigate scenes.

24. The active perception module according to claim 23, wherein said active perception module is implemented in a robot.

25. A robot comprising a decision making mechanism according to claim 1, wherein said robot is configured to autonomously investigate scenes.

26. A robot comprising an active perception module according to claim 23, wherein said robot is configured to autonomously investigate scenes.

22. A method for deciding on at least one prospective action of a robot from a set of actions of said robot, the method comprising the steps of:

computing a prior probabilistic representation of a prior environment state;

updating of said prior probabilistic representation with targets of a new observation on reducing at least one uncertainty in a posterior probabilistic representation of a posterior environment state to be reached after an appliance of said at least one prospective action, wherein said posterior probabilistic representation is a probabilistic representation resulting from said updating;

determining an information gain between said prior probabilistic representation and said posterior probabilistic representation by use of at least one information theoretic measure; and

evaluating said at least one prospective action by adding costs of executing said at least one prospective action to said information gain.