Massively parellel real-time network architectures for robots capable of self-calibrating their operating parameters through associative learning

• US 4,852,018 A
• Filed: 01/07/1987
• Issued: 07/25/1989
• Est. Priority Date: 01/07/1987
• Status: Expired due to Term

• Current Assignee: Trustees of Boston University
• Original Assignee: Trustees of Boston University
• Inventors: Kuperstein, Michael, Grossberg, Stephen
• Primary Examiner(s): RUGGIERO, JOSEPH F
• Application Number:
  US07/001,223
• Time in Patent Office: 930 Days
• Field of Search: 364/513, 364/413, 364/415, 364/417, 364/300, 364/200 MS File, 364/900 MS File, 364/413.02, 382/6, 382/14, 382/15, 128/731, 128/732, 128/733, 128/745, 351/203, 351/205, 351/211, 351/212, 351/224, 351/243, 351/246
• US Class Current: 700/259
• CPC Class Codes: B25J 9/163 learning, adaptive, model based, rule based expert control
• CPC Class Codes: B25J 9/1697 Vision controlled systems
• CPC Class Codes: G05B 19/414 Structure of the control system, e.g. common controller or multiprocessor systems, interface to servo, programmable interface controller
• CPC Class Codes: G05B 2219/33027 Artificial neural network controller
• CPC Class Codes: G05B 2219/42152 Learn, self, auto tuning, calibrating, environment adaptation, repetition
• 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: Y10S 706/904 Manufacturing or machine, e.g. agricultural machinery, machine tool

Abstract

A real-time network enables robots to accurately learn sensory motor transformation and to self-train and self-calibrate operating parameters after accidents or with wear. Combinations of visual and present position signals are used to relearn a target position map. Target positions in body-centered. visually activated coordinates are mapped into target positions in motor coordinates which are compared with present positions in motor coordinates to generate motor commands. Feedback provides calibrated error signals for adjustment of learned gain with changes in the system due to aging, accidents and the like. A series of prestored motor commands may be performed with a later "go" command.

Claims

1. A retinotopic command network for choosing a visual point of light for fixation from a field of visual points of light comprising:

a photo-detection means having a region for foveating light;

a body for supporting the photo-detection means;

motors for moving the photo-detection means to a position for foveating light;

a choice mechanism for transforming a broad array of lights collected by the photo-detection means into a localized activation of a sector map, wherein the localized activation of the sector map generates spatially coded signals for activating motors which move the photo-detection means to a position which foveates the chosen visual point;

a competitive choice means in communication with the photo-detection means and the sector map for converting a broadly distributed input pattern into a more narrowly focussed activity pattern;

a means for conditioning the spatially coded signals transmitted to the motors by a conditionable movement circuit which corrects a previous inaccurate movement signal, the conditionable movement circuit comprising;

a short term memory means for storing an internal marker indicative of the position of the chosen light on the photo-detection means before moving the photo-detection means; and

an error signal generated by the chosen light on the photo-detection means after the photo-detection has moved; and

a saccade generator which converts the conditioned spatially coded signals into a temporally coded signal that determines how long, and in what direction the motors will move the photo-detection means towards the chosen light.

2. A network for choosing a visual point of light for fixation from a field of visual points of light as claimed in claim 1, wherein the error signal is generated by comparing the location of the chosen light after the movement of the photo-detection means takes place with the position of the photo-detection mean which will foveate the chosen visual point.

3. A network for choosing a visual point of light for fixation from a field of visual points of light as claimed in claim 1 further comprising means for parallel processing of the spatially coded signals and error signals for conditioning movement signals transmitted to the motors for correcting previous errors which lead to an inaccurate movement of the photo-detection means.

4. A network for choosing a visual point of light for fixation from a field of visual points of light as claimed in claim 1 wherein the motors comprise three agonist-antagonist pairs of motors, wherein each motor is activated by a region of the section map.

5. A network for choosing a visual point of light for fixation from a field of visual points of light as claimed in claim 1 wherein the conditionable movement circuit is activated by explicit information signals generated by comparing a photo-detection map having coordinates which locate light detected by the photo-detection means with a photo-detection means position map having coordinates which locate the photo-detection means relative to the body supporting the detection means.

6. A network for choosing a visual point of light for fixation from a field of visual points of light as claimed in claim 1 wherein the conditionable movement pathway is activated by implicit information signals which are generated by a single invariant target position map having coordinates which locate the position of the chosen light on the photo-detection means for processing by the saccade generator.

7. A network for choosing a visual point of light for fixation from a field of visual points of light as claimed in claim 1 further comprising a head-muscle interface circuit means which mediates between target position coordinates computations and photo-detection coordinates computations to prevent new movement commands from being generated due to motion generated by the network choosing the visual point of light.

8. A network for choosing a visual point of light for fixation from a field of visual points of light as claimed in claim 1 wherein the sector map is subdivided into six equal sections, having a point in common with a region which foveates light, each of the six equal sections having a gradient of connections from each point in the section for communication with one of the motors such that the more eccentric points cause a more asymmetric movement of the light detection means.

9. A network for choosing a visual point of light for fixation from a field of visual points of light as claimed in claim 1 wherein a image stabilizer reflex means, which independently calculates the movement of the body and the movement of the photo-detection means to generate a total movement which compensates for the movement of the body, while maintaining the foveation of the chosen visual point.

10. A network for choosing a visual point of light for fixation from a field of visual points of light as claimed in claim 9 wherein the image stabilizing means are sufficiently slow to permit continuous registration of visual feedback to generate error signals.

11. A vector command network for computing vector differences comprising a head-muscle interface circuit which comprises a target position that is computed in head coordinates with a present photo-detection position computed in motor coordinates to generate a vector difference used to generate a movement signal for moving the photo-detection means to a position where the vector difference becomes zero, wherein the target position is encoded from head coordinates comprising a target position map into motor coordinates for comparison with the photo-detection position.

12. A vector command network for computing vector differences as claimed in claim 11 wherein the vector differences are encoded from motor coordinates into positional locations on the spatial map for making use of visually reactive movement commands that can be corrected using error signals.

13. A vector command network for computing vector differences as claimed in claim 11 wherein each target position in the target position map sends sampling signals to motor representations of all the motor position at the head-muscle interface.

14. A vector command network for computing vector differences as claimed in claim 11 wherein the position of the photo-detection means is determined by movement signals used to move and to hold the photo-detection means in position.

15. A vector command network for computing vector differences as claimed in claim 11 wherein an adaptive command system modifies the movement signals such that they are a linear function of the vector difference which gives rise to the movement signal.

16. A head-muscle interface circuit comprising;

a photo-detection means;

a means for receiving sampling signals over a conditionable circuit from an active population within a target position map;

a means for receiving corollary discharge signals which provide photo-detection position data that the conditionable circuit will learn for encoding the photo-detection position as a function of time such that the target position stored within the target position may can learn the photo-detection position that is attained by a subsequent movement of the photo-detection means;

wherein the conditionable pathways from the target position map to the head-muscle interface can learn only after a photo-detection movement is over; and

gating signals which is generated after a movement is over for modulating the learning that occurs within an adaptive weight system of active conditionable pathways.

17. A head-muscle interface circuit as claimed in claim 16 wherein the corollary discharge signals are in the form of a pattern of excitatory inputs for reading the position of the photo-detection means relative to a body supporting the photo-detection means.

18. A head-muscle interface circuit as claimed in claim 16 wherein the gating signals are activated after a saccade ends.

19. A vector command network for automatically compensating the difference between a present photo-detection position and a target photo-detection position comprising a head-muscle interface circuit which compares a target position that is computed in head coordinates with a present photo-detection position computed in motor coordinates to generate a vector difference used to generate a movement signal for moving the photo-detection means to a position where the vector difference becomes zero, wherein the target position is encoded from head coordinates comprising a target position map into motor coordinates for comparison with the photo-detection position.

20. A linearization network for linearizing the response of nonlinear motor means comprising:

inflow signals indicating the present position of the nonlinear motor means;

unconditional movement signals for initiating a movement signal for activating the nonlinear motor means;

an outflow-inflow stage which compares the unconditional movement signals with the inflow signals to generate error signals;

an adaptive gain stage which generates conditioned movement signals whose gain is dependent on the gain of the error signal and the unconditioned movement signal; and

a motorneuron stage for receiving the unconditioned movement signals and the conditioned movement signals to generate a total unconditioned movement signals which linearizes the response of the nonlinear motor means.

21. A linearization network for linearizing the movement of nonlinear motor means as claimed in claim 20, further comprising a head-muscle interface circuit which compares a target position with a present photo-detection means position determined by the unconditioned movement.

22. A linearization network for linearizing the movement of nonlinear motor means as claimed in claim 20, wherein the motorneurons are activated by tonic integrator cells which generate the unconditioned signals and medium lead burst cells which are target cells of a saccade generator which initiates a movement signal for moving the nonlinear motors.

23. A linearization network for linearizing the response of nonlinear motor means as claimed in claim 20 wherein the nonlinear motor means comprises at least one pair of nonlinear motors matched to form an agonist-antagonist pair, wherein competition between agonist and antagonist motor decreases the net antagonist output.

24. A linearization network for linearizing the movement of nonlinear motor means as claimed in claim 20, wherein the tonic cells generate corollary discharges to the head-muscle interface, excitatory signals to the outflow-inflow interface, and filter signals in communication with a photo-detection position map which samples error signals from the inflow-outflow interface to the adaptive gain stage.

25. A linearization network for linearizing the movement of nonlinear motor means as claimed in claim 20, wherein the outflow-inflow interface compensates for differences in the absolute size scales of the unconditional signals and the inflow signals by computing the relative sizes of the signals transmitted to the motors.

26. A linearization network for linearizing the movement of unlinear motor means as claimed in claim 20, further comprising an eye position map which recodes signals from the tonic cell, wherein the eye position map comprising a plurality of spatial locuses, each of which sends a separate sampling signal to the adaptive gain stage, wherein each separate sampling signal is adaptively encoded to reflect all gain changes that are caused by the error signal induced by the outflow-inflow interface.

27. A saccade generator circuit comprising:

a long-lead burster which adds signals from a retinotopic map which encodes a photo-detection movement and from a photo-detection position map which encodes the present position of a photo-detection means;

a medium-lead burster which receives the excitatory signals from the long-lead burster and first inhibitory signals which generate signals for activating tonic cells and motorneuron cells, wherein the tonic cells integrate their input signals and relay the integrated signals to the motorneuron cells which activates the motor for moving the photo-detection means;

pause cells which generate the first inhibitory signals received by the medium-lead burster; and

second inhibitory signals generated as a function of time by the long-head burster for inhibiting the pause cells from generating the first inhibitory signals;

wherein the tonic cells generate inhibitory feedback signals to the long-lead burster for terminating the movement of the photo-detection means.

28. A saccade generator circuit as claimed in claim 27 further comprising a postural circuit for compensating for any imbalances that may exist after a photo-detection movement is over to prevent the photo-detection means from drifting away from its target position.

29. A predictive command network that is capable of encoding a predictive sequence of movements in long term memory, and of reading out this movement sequence, comprising:

a first target position map for initially storing all relevant target positions in a short term memory means, wherein temporal order information of attended target positions is encoded by a spatial pattern of activation across the target positions with more intensely activated positions performed first;

a second target position map for selecting and storing the most active target positions stored within the first target position map for controlling a command movement of a photo-detection means; and

a head-muscle interface network in communication with the second target position map for generating output vectors, wherein the head-muscle interface network calculates matches and mismatches between target positions in the second target position map and present positions detected by the photo-detection means for regulating the sequential performance and learning of predictive movements.

30. A predictive command network as claimed in claim 29, wherein the chosen target position in the second target position map is stored in a first short term memory means until the movement corresponding to this chosen target is completed.

31. A predictive command network as claimed in claim 29, wherein a mismatch at the head-muscle interface generates a vector difference for generating an output pattern to a retinotopic map which stores the vector difference in a second short term memory means as a spatial location for generating signals to cause a photo-detection movement.

32. A predictive command network as claimed in claim 29, wherein a match at the head-muscle interface inhibits the stored target position in the second target position map from generating signals to the head-muscle interface which activates the next target position stored in the short term memory at the first target position map for storage in the second target position map.

33. A predictive command network for moving a photo-detection means to that a target light is foveated, comprising:

a first target position map for storing a spatial pattern of temporal command information which simultaneously stores several target lights to be foveated in short term memory;

a second target position map in communication with the first target position map for detecting and storing the most active target position detected in the first target position map in short term memory until the target light at that position is foveated; and

a head-muscle interface circuit having means for detecting the present position of the target light stored in the second target position map and comparing the present position of the target light detected with the target light stored to generate a movement signal for moving the photo-detection means during a mismatch and triggering a read-out signal of the next most active target position stored in the first target position map for storage in the second target position map.

34. A self-organizing input position map comprising:

a retinotopic map, which stores positions of light detected on a photo-detection means in a short term memory means before a movement begins;

a first photo-detection position map, which stores an initial photo-detection means position in short term memory before a photo-detection movement begins;

a retinotopic command network for improving the accuracy of the saccades;

a second photo-detection position map for encoding a target position after the retinotopic command network has improved the accuracy of a photo-detection movement; and

a now print gate means for allowing the retinotopic map and the first photo-detection position map to sample the locus of activity within the second photo-detection position map after each accurate movement of the photo-detection means.

35. A self-organizing target position map as claimed in claim 34 wherein the first photo-detection position map generates signals to a saccade generator for updating the position of the photo-detection means.

36. A self-organizing target position map as claimed in claim 34 wherein the retinotopic map generates signals for controlling direct unconditioned signals and indirect conditioned signals for generating visually reactive movements of the photo-detection means.

37. A system for coupling a retinotopic command network and a vector command network comprising:

a first retinotopic map for registering light sensitive activations on a photo-detection means;

a target position map in communication with the first retinotopic map having means activating a target position;

a head-muscle interface circuit for comparing the target position map with the present position of the photo-detection means position to generate vector difference used for activating a movement signal, which reposition the photo-detection means;

a second retinotopic map in communication with the head-muscle interface circuit for storing in short term memory a first light indicative of the target position until a photo-detection movement has terminated so that a second light, serving as an error signal, can be sampled;

a third retinotopic map in communication with the photo-detection means for registering light sensitive activations on the photo-detection means, and in communication with the second retinotopic maps for distributing conditionable signals across the second retinotopic map when the retinotopic command network generates correct foveations;

a long term memory trace which stores communication from the third retinotopic map to the second retinotopic map;

an adaptive gain stage for sampling unconditioned movement signals generated by the third retinotopic map and generating a conditioned movement signal;

a saccade generator for combining the unconditioned movement signals generated by the third retinotopic map and the conditioned movement signal generated by the adaptive gain stage for generating first movement signals;

tonic cells in communication with the saccade generator for generating corollary discharge signals to the head-muscle interface indicating the position of the photo-detection means; and

motorneurons in communication with the saccade generator and the tonic cells for combining the movement signals and corollary discharge signals to generate a total unconditional movement signal for moving the photo-detection means to a point which foveates the target.

38. A system for coupling a retinotopic command network and a vector command network comprising:

a first retinotopic map for recording light sensitive activations on a photo-detection means;

a first target position map in communication with the first retinotopic map having means for encoding several active target positions;

a second target position map in communication with the first target position map for choosing the target position from the first target position map which achieves the largest activation and the storing of that target position until the movement of the photo-detection means has terminated;

a target position map in communication with the first retinotopic map having means for receiving a spatial pattern of activated retinotopic positions such that a target position is activated;

a head-muscle interface circuit in communication with the first target position map for comparing the target position map with the present position of the photo-detection means position to generator vector differences used for activating the direction and length of a movement signal which repositions the photo-detection means;

a second retinotopic map in communication with the head-muscle interface circuit for storing in short term memory a first light indicative of the target position until movement of the photo-detection means has terminated so that a second light, serving as a visual error signal, can be sampled;

a third retinotopic map in communication with the photo-detection means for recording and for determining the most intense light detected on the photo-detection means for generating an error signal;

a fourth rectinotopic map in communication with the third retinotopic map having means for storing in short term memory the most intense light detected by the third retinotopic map until the resposition of the photo-detection means terminates, and having means for generating an unconditional signal;

a long term memory trace which stores communication from the fourth retinotopic map to the second retinotopic map;

an adaptive gain stage which samples unconditioned movement signals generated by the fourth retinotopic map and the error signal from the third retinotopic map for generating a conditioned movement signal;

a saccade generator for combining the unconditioned movement signals generated by the fourth retinotopic map and the conditioned movement signal generated by the adaptive gain stage for generating the movement signals used for repositioning the photo-detection means;

tonic cells in communication with the saccade generator for generating corollary discharge signals to the head-muscle interface indicating the position of the photo-detection means; and

motorneurons in communication with the saccade generator and the tonic cells for combining the movement signals and corollary discharge signals to generate a total unconditional movement signal for moving the photo-detection means to a point which foveates the target.

39. A system for coupling a retinotopic command network and a vector command network as claimed in claim 38 wherein the first retinotopic map is in communication with the third retinotopic map for enhancing the retinotopic position that was chosen within the third retinotopic map and the fourth retinotopic map is in communication with the first retinotopic map for receiving the entire spatial pattern of positions from the first retinotopic map.

40. A network for moving a photo-detection means relative to a body supporting the photo-detection means in response to light detected by the photo-detection means comprising:

a photo-detection position map for storing the position of the photo-detection means relative to the body;

a retinotopic map for storing light sensitive activations on the photo-detection means for generating a first unconditioned signal;

an adaptive gain stage in communication with the photo-detection position map and the retinotopic map for generating conditioned movement signals;

a saccade generator for combining the first unconditioned signal and the conditioned movement signal to generate a second unconditioned movement signal;

tonic cells in communication with the saccade generator for generating corollary discharge signals; and

motorneurons in communication with the saccade generator, the tonic cells, and the adaptive gain stage for combining the movement signals, conditioned movement signals and the corollary discharge signals to generate a total movement signal for moving the photo-detection means relative to the body.

41. A tension equalization network for preventing post-movement drift comprising:

a photo-detection means;

a body supporting the photo-detection means;

motor means responsive to movement signals for moving the photo-detection means relative to the body;

tonic cells for detecting activity patterns from a visual field on the photo-detection means;

a photo-detection position map for spatially encoding the activity patterns detected by the tonic cells for generating conditionable signals;

an accessory optic system for generating error signals when motions of the whole visual field are detected;

an adaptive gain stage in communication with the photo-detection position map and the accessory optic system for processing the conditionable signals and the error signals to generate the movement signals; and

means for generating a gating signal at the end of a movement of the photo-detection means relative to the body for activating the motor means so that it is receptive to the movement signals.