Legged mobile robot and its motion teaching method, and storage medium

• US 20040036437A1
• Filed: 07/21/2003
• Published: 02/26/2004
• Est. Priority Date: 04/03/2001
• Status: Active Grant

• Current Assignee: Sony Corporation (Sony Group Corp.)
• Original Assignee: Sony Corporation (Sony Group Corp.)
• Inventors: Ito, Masato
• Granted Patent: US 7,024,276 B2
• Time in Patent Office: Days
• US Class Current: 318/568.12
• 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

Learning-type motion control is performed using a hierarchical recurrent neural network. A motion pattern provided through human teaching work is automatically time-serially segmented with the hierarchical recurrent neural network, and the motion control of a machine body is carried out with a combination of the segmented data, whereby various motion patterns can be produced. With the time-serial segmentation, local time-serial patterns and an overall pattern as a combination of the local time-serial patterns are produced. For those motion patterns, indices for static stability and dynamic stability of the machine body, e.g., ZMP stability criteria, are satisfied and hence control stability is ensured.

Claims

1. A legged moving robot including drive units with plural degrees of freedom of joints, the robot comprising:

a motion pattern input unit for inputting a motion pattern from the exterior;

a plurality of determining units for previously storing local feasible motion patterns regarding respective degrees of freedom of joints, and each determining agreement between a motion pattern inputted from said motion pattern input unit and the stored motion pattern;

a gate unit for gating a determination result in each of said determining units;

a merging unit for merging outputs from said gate unit and producing an overall motion pattern for a machine body of said legged moving robot; and

a taught motion pattern realizing unit for storing a totalized result in said merging unit as taught motion pattern data, and/or reproducing the result on the machine body of said legged moving robot.

2. A legged moving robot according to claim 1, wherein said motion pattern input unit is constituted as a motion capturing device for capturing, by a camera, measurement points attached to various positions over the whole body of an operator, tracking movements of the measurement points, and obtaining motions of a skeleton to realize the taught motion pattern, or constituted as a device for inputting motion pattern data obtained through said motion capturing device.

3. A legged moving robot according to claim 1, wherein said motion pattern input unit is constituted as a device for inputting motion pattern data obtained through a teaching pendant.

4. A legged moving robot according to claim 1, wherein said motion pattern input unit is constituted as a device for inputting motion pattern data obtained through direct teaching directly made on the machine body of said legged moving robot.

5. A legged moving robot according to claim 1, wherein the feasible motion pattern is a motion pattern for which kinematic matching and/or dynamic matching is satisfied in the machine body of said legged moving robot.

6. A legged moving robot according to claim 1, wherein each of said plurality of determining units is constituted as a recurrent neural network comprising an input layer constituted as a unit group to which input data is inputted, an output layer constituted as a unit group which outputs a network output, and an intermediate layer constituted as a unit group other than said unit groups,

said input layer being made up of an input unit to which a local time-serial motion pattern for the machine body of said legged moving robot is inputted, and a context unit,

said intermediate layer being constituted as a hidden unit,

said output layer being made up of an output unit which outputs a prediction of the local time-serial motion pattern, and a context unit,

a part of the output from said output layer being fed back to the context unit of the input layer via a context loop, whereby learning of the time-serial motion pattern is performed.

7. A legged moving robot according to claim 1, wherein said gate unit is constituted as a recurrent neural network comprising an input layer constituted as a unit group to which input data is inputted, an output layer constituted as a unit group which outputs a network output, and an intermediate layer constituted as a unit group other than said unit groups,

said input layer being made up of an input unit to which a gate serving as a weight applied to the determining unit is inputted, and a context unit,

said intermediate layer being constituted as a hidden unit,

said output layer being made up of an output unit which outputs a gate serving as a weight applied to the determining unit, and a context unit,

a part of the output from said output layer being fed back to the context unit of the input layer via a context loop, whereby learning of the time-serial gate is performed.

8. A method of teaching motions in a legged moving robot including drive units with plural degrees of freedom of joints, the method comprising the steps of:

a motion pattern inputting step of inputting a motion pattern from the exterior;

a plurality of determining steps of previously storing local feasible motion patterns regarding respective degrees of freedom of joints, and each determining agreement between a motion pattern inputted in said motion pattern inputting step and the stored motion pattern;

a gating step of gating a determination result in each of said determining steps;

a merging step of merging outputs from said gating step and producing an overall motion pattern for a machine body of said legged moving robot; and

a taught motion pattern realizing step of storing a totalized result in said merging step as taught motion pattern data, and/or reproducing the result on the machine body of said legged moving robot.

9. A method of teaching motions in a legged moving robot according to claim 8, wherein said motion pattern inputting step is constituted by a device for inputting motion pattern data obtained through a motion capture, said motion capture capturing, by a camera, measurement points attached to various positions over the whole body of an operator, tracking movements of the measurement points, and obtaining motions of a skeleton to realize the taught motion pattern.

10. A method of teaching motions in a legged moving robot according to claim 8, wherein said motion pattern inputting step inputs motion pattern data obtained through a teaching pendant.

11. A method of teaching motions in a legged moving robot according to claim 8, wherein said motion pattern inputting step inputs motion pattern data obtained through direct teaching directly made on the machine body of said legged moving robot.

12. A method of teaching motions in a legged moving robot according to claim 8, wherein the feasible motion pattern is a motion pattern for which kinematic matching and/or dynamic matching is satisfied in the machine body of said legged moving robot.

13. A method of teaching motions in a legged moving robot according to claim 8, wherein each of said plurality of determining step is constituted by a recurrent neural network comprising an input layer constituted as a unit group to which input data is inputted, an output layer constituted as a unit group which outputs a network output, and an intermediate layer constituted as a unit group other than said unit groups,

said input layer being made up of an input unit to which a local time-serial motion pattern for the machine body of said legged moving robot is inputted, and a context unit,

said intermediate layer being constituted as a hidden unit,

said output layer being made up of an output unit which outputs a prediction of the local time-serial motion pattern, and a context unit,

a part of the output from said output layer being fed back to the context unit of the input layer via a context loop, whereby learning of the time-serial motion pattern is performed.

14. A method of teaching motions in a legged moving robot according to claim 8, wherein said gating step is constituted as a recurrent neural network comprising an input layer constituted as a unit group to which input data is inputted, an output layer constituted as a unit group which outputs a network output, and an intermediate layer constituted as a unit group other than said unit groups,

said input layer being made up of an input unit to which a gate serving as a weight applied to the determining unit is inputted, and a context unit,

said intermediate layer being constituted as a hidden unit,

said output layer being made up of an output unit which outputs a gate serving as a weight applied to the determining unit, and a context unit,

a part of the output from said output layer being fed back to the context unit of the input layer via a context loop, whereby learning of the time-serial gate is performed.

15. A storage medium product physically storing, in a computer readable form, computer software described to execute, on a computer system, a process of teaching motions in a legged moving robot including drive units with plural degrees of freedom of joints, the computer software comprising the steps of:

a motion pattern inputting step of inputting a motion pattern from the exterior;

a plurality of determining steps of previously storing local feasible motion patterns regarding respective degrees of freedom of joints, and each determining agreement between a motion pattern inputted in said motion pattern inputting step and the stored motion pattern;

a gating step of gating a determination result in each of said determining steps;

a merging step of merging outputs from said gating step and producing an overall motion pattern for a machine body of said legged moving robot; and

a taught motion pattern realizing step of storing a totalized result in said merging step as taught motion pattern data, and/or reproducing the result on the machine body of said legged moving robot.