Machine learning approach for predicting humanoid robot fall

• US 8,554,370 B2
• Filed: 01/29/2010
• Issued: 10/08/2013
• Est. Priority Date: 05/15/2009
• Status: Active Grant

• Current Assignee: Honda Motor Co., Ltd. (Honda Motor Company)
• Original Assignee: Honda Motor Co., Ltd. (Honda Motor Company)
• Inventors: Goswami, Ambarish, Kalyanakrishnan, Shivaram
• Primary Examiner(s): Tran, Khoi
• Assistant Examiner(s): MOYER, DALE S
• Application Number:
  US12/696,783
• Publication Number:
  US 20100292838A1
• Time in Patent Office: 1,348 Days
• Field of Search: 700245-264, 901 1- 52
• US Class Current: 700/253
• CPC Class Codes: B25J 9/163 learning, adaptive, model based, rule based expert control
• CPC Class Codes: B62D 57/032 with alternately or sequentially lifted supporting base and legs; with alternately or sequentially lifted feet or skid B62D57/024 takes precedence
• 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

Abstract

A system and method is disclosed for predicting a fall of a robot having at least two legs. A learned representation, such as a decision list, generated by a supervised learning algorithm is received. This learned representation may have been generated based on trajectories of a simulated robot when various forces are applied to the simulated robot. The learned representation takes as inputs a plurality of features of the robot and outputs a classification indicating whether the current state of the robot is balanced or falling. A plurality of features of the current state of the robot, such as the height of the center of mass of the robot, are determined based on current values of a joint angle or joint velocity of the robot. The current state of the robot is classified as being either balanced or falling by evaluating the learned representation with the plurality of features of the current state of the robot.

Claims

1. A computer-implemented method for predicting a fall of a robot having at least two legs, the method comprising:

generating a learned representation using a supervised learning algorithm, the learned representation taking as inputs a plurality of features of a robot, the learned representation having as an output a classification comprising an indication of a balanced state or a falling state, wherein generating the learned representation using the supervised learning algorithm comprises steps of;

applying a plurality of simulated force impulses to a simulation of the robot, the force impulses varying in magnitude of force and direction of application;

recording a plurality of trajectories generated from the motions of the robot after application of the plurality of simulated force impulses, each trajectory comprising a plurality of instances, each instance comprising a plurality of features describing the state of the robot at the particular instance;

classifying, by a processing device, each instance as a balanced instance or as a falling instance based on whether the trajectory containing the instance ends in a fallen state; and

processing the features and classification of each instance with the supervised learning algorithm to generate the learned representation;

determining a plurality of features of a current state of the robot, the determining based at least in part on a current value of a joint angle or joint velocity of the robot; and

classifying the current state of the robot as being either balanced or falling, the classifying performed by evaluating the learned representation with the plurality of features of the current state of the robot.

2. The method of claim 1, wherein generating the learned representation using the supervised learning algorithm further comprises:

removing an instance classified as falling if the instance occurs before a specific time in a trajectory containing the instance, wherein the specific time is based on a time of onset of a monotonic decrease of a height of a center of mass of the robot.

3. The method of claim 1, wherein the learned representation comprises a decision list.

4. The method of claim 1, further comprising:

responsive to classifying the current state of the robot as being falling, activating a fall controller of the robot.

5. The method of claim 4, wherein the fall controller of the robot controls the robot to fall in a manner that reduces damage to the robot or an object near the robot.

6. The method of claim 1, wherein one of the plurality of features comprises a height of the center of mass of the robot.

7. The method of claim 1, wherein one of the plurality of features comprises a foot contact mode of the robot.

8. The method of claim 7, wherein the learned representation comprises a plurality of decision lists each associated with a different foot contact mode of the robot.

9. A non-transitory computer-readable storage medium storing executable computer program modules for predicting a fall of a robot having at least two legs, the computer program modules when executed by a processor causing the processor to perform steps including:

generating a learned representation using a supervised learning algorithm, the learned representation taking as inputs a plurality of features of a robot, the learned representation having as an output a classification comprising an indication of a balanced state or a falling state, wherein generating the learned representation using the supervised learning algorithm further comprises steps of;

applying a plurality of simulated force impulses to a simulation of the robot, the force impulses varying in magnitude of force and direction of application;

recording a plurality of trajectories generated from the motions of the robot after application of the plurality of simulated force impulses, each trajectory comprising a plurality of instances, each instance comprising a plurality of features describing the state of the robot at the particular instance;

classifying each instance as a balanced instance or as a falling instance based on whether the trajectory containing the instance ends in a fallen state; and

processing the features and classification of each instance with the supervised learning algorithm to generate the learned representation;

determining a plurality of features of a current state of the robot, the determining based at least in part on a current value of a joint angle or joint velocity of the robot; and

classifying the current state of the robot as being either balanced or falling, the classifying performed by evaluating the learned representation with the plurality of features of the current state of the robot.

10. The non-transitory computer-readable storage medium of claim 9, wherein generating the learned representation using the supervised learning algorithm further comprises:

removing an instance classified as falling if the instance occurs before a specific time in a trajectory containing the instance, wherein the specific time is based on a time of onset of a monotonic decrease of a height of a center of mass of the robot.

11. The non-transitory computer-readable storage medium of claim 9, wherein the learned representation comprises a decision list.

12. The non-transitory computer-readable storage medium of claim 9, wherein the modules are further configured for:

responsive to classifying the current state of the robot as being falling, activating a fall controller of the robot.

13. The non-transitory computer-readable storage medium of claim 12, wherein the fall controller of the robot controls the robot to fall in a manner that reduces damage to the robot or an object near the robot.

14. The non-transitory computer-readable storage medium of claim 9, wherein one of the plurality of features comprises a height of the center of mass of the robot.

15. The non-transitory computer-readable storage medium of claim 9, wherein one of the plurality of features comprises a foot contact mode of the robot.

16. The non-transitory computer-readable storage medium of claim 15, wherein the learned representation comprises a plurality of decision lists each associated with a different foot contact mode of the robot.