Machine learning device, robot control system, and machine learning method
A machine learning device that acquires state information from a robot control inspection system. The system has a robot hand to hold a workpiece or camera. The state information includes a flaw detection position of the workpiece, a movement route of the robot hand, an imaging point of the workpiece, and the number of imaging by the camera. A reward calculator calculates a reward value in reinforcement learning based on flaw detection information including the flaw detection position. A value function updater updates an action value function by performing the reinforcement learning based on the reward value, the state information, and the action.
1. A machine learning device for use with, and configured to perform reinforcement learning with respect to a robot control system,
the robot control system comprising:
an illumination means that irradiates a surface to be inspected of an object to be inspected with illumination light;
an imaging means that images the surface to be inspected;
a robot that includes a robot hand;
a control unit that, while moving the robot hand gripping the object to be inspected or the imaging means, along a movement route including a plurality of imaging points set on the surface to be inspected so that the surface to be inspected is entirely covered by a plurality of images imaged by the imaging means, causes the imaging means to image the imaging points set on the surface to be inspected; and
a flaw inspection unit that detects a flaw on the surface to be inspected on the basis of the image obtained by imaging the surface to be inspected by the imaging means,
the machine learning device comprising:
an action information output unit that outputs action information including adjustment information of the imaging region including the imaging points, to the control unit;
a state information acquisition unit that acquires state information from the control unit and the flaw inspection unit resulting from a number N of images, obtained by imaging the surface to be inspected by the imaging means by moving the robot hand gripping the object to be inspected or the imaging means by the control unit, based on the action information, the state information including the number N and flaw detection information, the flaw detection information including a flaw detection position of the surface to be inspected, detected by the flaw inspection unit, with respect to each of a plurality of objects to be inspected prepared in advance;
a reward output unit that outputs a reward value in the reinforcement learning based on the number N and the flaw detection information including the flaw detection positions included in the state information; and
a value function updating unit that updates an action value function based on the reward value, the state information, and the action information.
2. The machine learning device according to claim 1 , wherein the reward output unit
sets the reward value to be a positive value when: (i) the number of the flaw detection positions detected by the flaw inspection unit is equal to the number of the flaw detection positions detected by the flaw inspection unit in an initial state at machine learning start time of imaging a imaging region including predetermined imaging points, and (ii) the number N is smaller than the number of images in the initial state of the machine learning start time, and
sets the reward value to be a negative value when the number of the flaw detection positions detected by the flaw inspection unit is smaller than the number of the flaw detection positions detected by the flaw inspection unit in the initial state of the machine learning start time of imaging the imaging region including the predetermined imaging points.
3. The machine learning device according to claim 2 , wherein the reward output unit
sets the reward value to be a larger value as the number N is smaller than the number of images in the initial state of the machine learning start time, when the number of the flaw detection positions detected by the flaw inspection unit is equal to the number of the flaw detection positions detected by the flaw inspection unit in the initial state of the machine learning start time of imaging the imaging region including the predetermined imaging points.
4. The machine learning device according to claim 2 , comprising an optimizing action information output unit that generates optimizing action information with which the number of the flaw detection positions is equal to the number of the flaw detection positions detected by the flaw inspection unit in the initial state of the machine learning start time, and the number N is a minimum, based on the action value function updated by the value function updating unit.
5. The machine learning device according to claim 1 , wherein the reward output unit
sets the reward value to be a positive value when: (i) the number of the flaw detection positions in the predetermined region of the surface to be inspected, detected by the flaw inspection unit is equal to the number of the flaw detection positions in the predetermined region of the surface to be inspected, detected by the flaw inspection unit in the initial state of the machine learning start time of imaging the imaging region including the predetermined imaging points, and (ii) the number N is smaller than the number of images in the initial state of the machine learning start time,
and sets the reward value to be a negative value when the number of the flaw detection positions in the predetermined region of the surface to be inspected, detected by the flaw inspection unit is smaller than the number of the flaw detection positions in the predetermined region of the surface to be inspected, detected by the flaw inspection unit in the initial state of the machine learning start time of imaging the imaging region including the predetermined imaging points.
6. The machine learning device according to claim 5 , wherein the reward output unit
sets the reward value to be a larger value as the number N is smaller than the number of images in the initial state of the machine learning start time, when the number of the flaw detection positions in the predetermined region of the surface to be inspected, detected by the flaw inspection unit is equal to the number of the flaw detection positions in the predetermined region of the surface to be inspected, detected by the flaw inspection unit in the initial state of the machine learning start time of imaging the imaging region including the predetermined imaging points.
7. The machine learning device according to claim 5 , comprising an optimizing action information output unit that generates optimizing action information with which the number of the flaw detection positions in the predetermined region of the surface to be inspected is equal to the number of the flaw detection positions in the predetermined region of the surface to be inspected, detected by the flaw inspection unit in the initial state at the machine learning time by imaging the imaging region including the predetermined imaging points, and the number N is a minimum, based on action value function updated by the value function updating unit.
8. The machine learning device according to claim 1 , wherein
the action value function is shared with other machine learning devices, and
the value function updating unit updates the action value function that is shared.
9. A system comprising the machine learning device and the robot control system of claim 1 .
10. A machine learning method by a machine learning device that performs reinforcement learning with respect to a robot control system,
the robot control system comprising:
an illumination means that irradiates a surface to be inspected of an object to be inspected with illumination light;
an imaging means that images the surface to be inspected;
a robot that includes a robot hand;
a control unit that, while moving the robot hand gripping the object to be inspected or the imaging means, along a movement route including a plurality of imaging points set on the surface to be inspected so that the surface to be inspected is entirely covered by a plurality of images imaged by the imaging means, causes the imaging means to image the imaging points set on the surface to be inspected; and
a flaw inspection unit that detects a flaw on the surface to be inspected on the basis of the image obtained by imaging the surface to be inspected by the imaging means,
the machine learning method comprising:
an action information output step of outputting action information including adjustment information of the imaging region including the imaging points, to the control unit;
a state information acquisition step of, with respect to each of a plurality of objects to be inspected prepared in advance, moving the robot hand gripping the object to be inspected or the imaging means by the control unit on the basis of the action information, and acquiring state information from the control unit and the flaw inspection unit resulting from a number N of images, obtained by imaging the surface to be inspected by the imaging means by moving the robot hand gripping the object to be inspected or the imaging means by the control unit, based on the action information, the state information including the number N and flaw detection information, the flaw detection information including a flaw detection position detected by the flaw inspection unit, and the number of imaging, on the basis of an image obtained by imaging the surface to be inspected by the imaging means;
a reward output step of outputting a reward value in the reinforcement learning based on the number N and the flaw detection information including the flaw detection positions included in the state information; and
a value function updating step of updating an action value function based on the reward value, the state information, and the action information.