Apparatus and methods for training of robots

• US 9,687,984 B2
• Filed: 12/31/2014
• Issued: 06/27/2017
• Est. Priority Date: 10/02/2014
• Status: Active Grant

• Current Assignee: Brain Corporation
• Original Assignee: Brain Corporation
• Inventors: Smith, Andrew T., Polonichko, Vadim
• Primary Examiner(s): Kiswanto, Nicholas
• Assistant Examiner(s): Bendidi, Rachid
• Application Number:
  US14/588,168
• Publication Number:
  US 20160096272A1
• Time in Patent Office: 909 Days
• Field of Search: 700245, 700253, 700255, 901 9
• CPC Class Codes: B25J 9/0081 with master teach-in means
• CPC Class Codes: B25J 9/1602 characterised by the control system, structure, architecture
• CPC Class Codes: B25J 9/1607 Calculation of inertia, jacobian matrixes and inverses
• CPC Class Codes: B25J 9/161 Hardware, e.g. neural networks, fuzzy logic, interfaces, processor
• CPC Class Codes: B25J 9/163 learning, adaptive, model based, rule based expert control
• CPC Class Codes: B25J 9/1666 Avoiding collision or forbidden zones
• CPC Class Codes: B25J 9/1697 Vision controlled systems
• CPC Class Codes: G05D 1/0088 characterized by the autonomous decision making process, e.g. artificial intelligence, predefined behaviours using knowledge based models G06N5/00
• CPC Class Codes: G05D 1/0246 using a video camera in combination with image processing means
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 3/00 Computing arrangements based on biological models
• CPC Class Codes: G06N 3/008 based on physical entities controlled by simulated intelligence so as to replicate intelligent life forms, e.g. based on robots replicating pets or humans in their appearance or behaviour
• CPC Class Codes: G06N 3/049 Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
• CPC Class Codes: Y10S 901/01 Mobile robot
• CPC Class Codes: Y10S 901/03 Teaching system
• CPC Class Codes: Y10S 901/09 Closed loop, sensor feedback controls arm movement
• CPC Class Codes: Y10S 901/47 Optical

Abstract

A random k-nearest neighbors (RKNN) approach may be used for regression/classification model wherein the input includes the k closest training examples in the feature space. The RKNN process may utilize video images as input in order to predict motor command for controlling navigation of a robot. In some implementations of robotic vision based navigation, the input space may be highly dimensional and highly redundant. When visual inputs are augmented with data of another modality that is characterized by fewer dimensions (e.g., audio), the visual data may overwhelm lower-dimension data. The RKNN process may partition available data into subsets comprising a given number of samples from the lower-dimension data. Outputs associated with individual subsets may be combined (e.g., averaged). Selection of number of neighbors, subset size and/or number of subsets may be used to trade-off between speed and accuracy of the prediction.

Claims

1. A method of determining a control signal for a robot based on a random k-nearest neighbors (RKNN) learning process, the method being performed by a computing platform having one or more processors executing instructions stored by a non-transitory computer-readable storage medium, the method comprising:

receiving first input features of a first type and second input features of a second type;

determining a subset of features comprising a portion of the first input features and at least one feature from the second input features, where the determining further comprises dynamically selecting a number of features in the subset based on a speed of computation and an accuracy of a prediction;

comparing individual features of the subset to corresponding features of a plurality of training feature sets;

based on the comparison, determining a similarity measure for a given training set of the plurality of training feature sets, the similarity measure characterizing a similarity between the individual features of the subset and features of the given training set;

responsive to the similarity measure breaching a threshold, selecting one or more training sets from the plurality of training feature sets;

determining one or more potential control signals for the robot, individual ones of the one or more potential control signals being associated with a corresponding training set of the plurality of training feature sets; and

determining the control signal based on a RKNN transformation obtained from the one or more potential control signals;

wherein the control signal is configured to cause the robot to execute an action.

2. The method of claim 1, wherein the similarity measure is determined based on a difference between each first value of the features of the subset and each second value of the features of the given training set.

3. The method of claim 1, wherein the similarity measure is determined based on a distance metric between the individual features of the subset of features and corresponding features of the given training set.

4. The method of claim 3, wherein the selecting of the one or more training sets comprises selecting a training set associated with a smallest distance metric.

5. The method of claim 3, wherein the selecting of the one or more training sets comprises selecting N training sets associated with a lowest percentile of the distance metric, N being greater than two.

6. The method of claim 5, wherein the transformation comprises a statistical operation performed on individual ones of the one or more potential control signals associated with the selected N training sets.

7. The method of claim 6, wherein the statistical operation is selected from the group consisting of a mean and a percentile.

8. The method of claim 5, wherein the transformation comprises a weighted sum of a product of individual ones of the one or more potential control signals and a corresponding distance measure associated with the selected N training sets.

9. The method of claim 1, wherein:

the first input type comprises a digital image comprising a plurality of pixel values; and

the second input type comprises a binary indication associated with the action being executed.

10. The method of claim 9,

further comprising training the robot with a plurality of iterations configured based on a training signal from the robot;

wherein a given iteration is characterized by a control command and a performance measure associated with the action execution based on the control command.

11. The method of claim 9, wherein:

the plurality of pixels comprises at least 10 pixels; and

the determining of the subset of features further comprises performing a random selection based on a random number generation operation.

12. A self-contained robotic apparatus, the apparatus comprising:

a platform comprising a motor;

a first sensor component configured to provide a signal conveying a video frame comprising a plurality of pixels;

a second sensor component configured to provide a binary sensor signal characterized by one of two states;

a memory component configured to store training sets, a given training set comprising an instance of the video frame, an instance of the binary sensor signal, and an instance of a motor control indication configured to cause the robot to execute an action;

one or more processors configured to operate a random k-nearest neighbors learning process to determine a motor control indication by at least;

determining a subset of features comprising the binary sensor signal and a set of pixels randomly selected from the plurality of pixels;

scaling individual pixels of the set of pixels by a scaling factor;

scaling features of the subset of features by a scaling factor;

comparing individual scaled features of the subset to corresponding features of individual ones of the training sets;

based on the comparison, determining a similarity measure for a given training set, the similarity measure characterizing a similarity between features of the subset and features of the given training set;

based on an evaluation of the similarity measure, selecting one or more of the training sets;

determining one or more potential control signals for the robot, individual ones of the one or more potential control signals being associated with a corresponding training set; and

determining the control signal based on a RKNN transformation obtained from the one or more potential control signals.

13. The apparatus of claim 12, wherein:

the scaling comprises a multiplication of a first input by the scaling factor; and

the scaling factor is determined based on a number of pixels in the subset.

14. The apparatus of claim 13, wherein the scaling factor is determined based on a ratio of a range of pixel values to a range of the binary values.

15. The apparatus of claim 12, wherein:

the action comprises target-approach-obstacle-avoidance; and

the scaling is performed based on a size of obstacle or object as the obstacle or the object appears in the video frame.

16. The apparatus of claim 12, wherein the scaling is pixel specific.

17. A non-transitory computer-readable storage medium having instructions embodied thereon, the instructions being executable by a processor to perform a method of selecting an outcome of a plurality of outcomes according to a random k-nearest neighbors learning process, the method comprising:

determining a history of sensory input;

applying a transformation to an instance of the sensory input, the transformation configured, based on an analysis of the history, to produce a scaled input;

determining a set of features comprising features of a first type randomly selected from the scaled input and at least one feature of a second type;

comparing individual features of the set of features to corresponding features of a plurality of training feature sets, individual ones of the plurality of training feature sets comprising a number of training features, the number being equal to or greater than a quantity of features within the set of features;

based on the comparison, determining a similarity measure for a given training set of the plurality of training feature sets, the similarity measure characterizing a similarity between features of the subset and features of the given training set;

responsive to the similarity measure breaching a threshold, selecting one or more training sets from the plurality of training sets;

determining one or more potential control signals for a robot, individual ones of the one or more potential control signals being associated with a corresponding training set of the plurality of training sets; and

determining a control signal based on a RKNN transformation obtained from the one or more potential control signals, the control signal being configured to cause the robot to execute a physical task.

18. The storage medium of claim 17, wherein:

the analysis of the history comprises a determination of a feature mean and a feature standard deviation; and

the transformation to the instance of the sensory input comprises subtracting the feature mean and dividing an outcome by the feature standard deviation.

19. The storage medium of claim 17, wherein:

the set comprises a plurality of set features, individual ones of the set features characterized by a pointer; and

the feature mean and the feature standard deviation are configured for a respective pointer.

20. The storage medium of claim 19, wherein:

the input of the first type comprises a matrix of values;

the pointer identifies a value within the matrix; and

the feature mean and the feature standard deviation are configured for a given location within the matrix.