CN113814989A

CN113814989A - Deformable combined robot and control system thereof

Info

Publication number: CN113814989A
Application number: CN202010563621.6A
Authority: CN
Inventors: 周世海
Original assignee: Individual
Current assignee: Individual
Priority date: 2020-06-19
Filing date: 2020-06-19
Publication date: 2021-12-21

Abstract

The invention discloses a deformable combined robot and a control system thereof, which comprise N cell units, wherein the N cell units are in closable communication connection, N is a positive integer, and the arrangement can realize the variable interconnection of data channels among the cell units of the robot, so that the robot can better change an internal communication connection mode according to the environment condition, realize the adaptation to the environment and further develop corresponding function execution tasks.

Description

Deformable combined robot and control system thereof

Technical Field

The invention relates to the technical field of robots, in particular to a deformable combined robot, a system and a control method thereof.

Background

Among movie terminators, directors provide a liquid metal-based robot concept for people, and regarding the robot, the current research progress of the academic world only stays in the preliminary exploration of liquid metal programming, such as liquid metal software machines of the university of Qinghua, programmable water droplet controllers of the labor of Massachusetts, and the like. However, this is still a long way to go compared to the liquid metal end-up in a movie.

Although the deformable robot of pure liquid metal does not work well, the deformable robot using other materials has corresponding material for support, such as a material that can move by itself only by illumination, as developed by research teams at eindhoven university in the netherlands and at kent state university in the united states. Scientists clamp this polymer material in a rectangular frame that, when illuminated with light, moves at the rate of caterpillar peristalsis, but does not require any impulse. The principle is that a fast-response photosensitive variant in a liquid crystal polymer network is integrated, one side of a material reflects light when the material is illuminated, and the other side expands, so that the material deforms and bulges when the material is illuminated, and the material immediately relaxes when light disappears. This particular ability makes it the world's first material to convert illumination directly into displacement.

The above material has characteristics very similar to the movement pattern of amoebae, which is a part of the body growing in the direction of movement and then having cytoplasm flowing therein. Obviously, the robot assembly similar to the cell can be manufactured by using the material similar to the material and capable of converting other energy into displacement, the assembly is controlled under the control of a machine learning algorithm, a low-match terminator capable of learning skills and changing the form of the robot according to task requirements is obtained, and the robot can change the form of the robot according to the environment and play a great role in the environments unsuitable for human presence, such as severe natural environments, dangerous battlefields and the like. However, the structural design and control algorithm of the robot, especially the control algorithm, are not disclosed in the academic world, so that the robot has sufficient material base for support, but still has no possibility of appearing, and the core problem is that the algorithm needs to be capable of realizing multi-thread tasks and has adaptability to the environment, which puts very high requirements on the machine learning algorithm, supervised learning needs a large amount of data for learning, and unsupervised learning needs a large amount of cost in learning, such as GPT 3.

However, as to how to design the algorithm of the robot, the academic world already has a relatively complete basic theory, and only lacks of an integrated guiding idea, so that the machine learning algorithm at the present stage still cannot obtain a significant breakthrough.

To provide guiding theory, the great britain neurologist karr frieston proposed a theory of unified significance that attempted to answer the question "why life would produce negative entropy" proposed by schrodinger in a principle known as the "free energy principle".

The second law of thermodynamics, entropy increase law, tells us that universe tends to have entropy increasing continuously and finally goes to chaotic death, entropy is disorder degree, everything tends to go from ordered to disordered, Schrodinger proposes that life exists positive entropy which needs to be offset continuously, and grows from negative entropy, and then a problem is generated why life generates negative entropy? We wake up every morning, or yesterday's me, how do the biological system resist the law of entropy increase?

The free energy principle of frieston states that all life is driven by the same general command, which can be reduced to a mathematical function, on any scale of the tissue, from single cells to the human brain. Alive is to reduce the difference between your desire and sensory input, i.e. to minimize free energy.

In short, the core of the above theory is that even if the cell scale is fine, there is a data processor and a predictor for the data appearing in the next cycle, and the free energy reduction means that when the actual data of the next cycle appears, the data is the same as the data obtained by the predictor, and the free energy is the lowest.

To reduce the free energy, further, carl proposes two ways to reduce the free energy, one to make the prediction practical by perceptual learning, and one to make active inference, i.e., to make the practice predictive by taking behavioral actions.

However, the carr's theory is not well consistent with the fact that human beings sometimes feel interested in something beyond accident and feel happy when driving by behavior, and then take actions to further enhance the happy, which motivation is not solved by carr's theory. Karl's theory can only explain the behavioral motivation of the cell itself, but not of human beings with emotions. To solve this problem, the applicant proposed an emotion-based reinforcement learning control method in patent application No. 2019112884985, which has a certain similarity with the free energy theory but a great difference in the concept behind, and which also has perception learning and active inference, but the perception learning of which is a knowledge-accepting process, and the active inference of which is a process of interest in a local feature and producing pursuit, and which is more adaptive to the human conscious emotion itself than the free energy theory, and in the prior art, an AI having perception learning ability and active inference ability has been designed based on the free energy theory and has a feature of exploring the world first and then performing a task in behavioral expression, and obviously which also has active inference ability and perception learning ability, and will have emotion in actual behavior than the AI, an AI with an emotion will on the one hand be more complex in its behavioral pattern than an AI without emotion, but on the other hand, due to the observable change in the point of execution of the AI, the behavioral intent of the AI will also be visualized and thus can be made more controllable by manual intervention. And when the method is applied to AI evolution, more directions and possibilities are provided.

In addition, at present, the AutoML is one of the more important directions in the AI field, a network model structure suitable for a specified task is obtained by combining, screening and mutating neural network components, the mechanism is actually combined from top to bottom, a neural network structure which has a single flow direction and can only execute a local task is obtained, an AI which can execute various tasks and has strong generalization capability cannot be generated well.

However, based on the above theory, how to design a deformable combined robot and a corresponding control method becomes a problem which is easier to solve

Disclosure of Invention

The invention aims to provide a deformable combined robot and a control system thereof, wherein the deformable combined robot adapts to the environment by deforming according to the environment and task requirements and has self-evolution capability.

The technical scheme adopted by the robot is that the deformable combined robot comprises N cell units, the N cell units are in closable communication connection, N is a positive integer, the cell units of the robot can be movably interconnected through a data channel, the robot can change an internal communication connection mode according to the environment condition better, the environment can be adapted, and corresponding function execution tasks can be developed.

Preferably, the robot further comprises a main controller, wherein the main controller is in communication connection with at least one cell unit in the N cell units, and the cell units are controlled through the main controller, so that the main controller can make a task target from the whole world and determine the evolution direction for the robot.

Preferably, the cell unit comprises a control chip, a shaping supporting layer and a movable layer connected to the shaping supporting layer, the control chip is in signal communication connection with the shaping supporting layer and the movable layer respectively, and the cell unit is provided with moving capacity and deformation capacity by the aid of the cell unit.

Preferably, the control chip at least comprises a distance sensor, a movable layer motion change decision unit, a shaping support layer shaping change control unit, a data processing unit for processing data interactive with other cell units, a data prediction unit for predicting data output by the data processing unit in the next period, and a behavior unit for converting transmission channels of the cell units and other cell units, wherein the distance sensor is respectively in signal connection with the motion change decision unit and the behavior unit; the shaping change decision unit is in signal connection with the data processing unit, and the data processing unit is in data connection with the data prediction unit; the data processing unit and the data prediction unit are in data connection with the behavior unit, and the arrangement enables the control chip to control the movement of the cell unit, so that the robot capable of adapting to the environment is combined.

The technical scheme adopted by the control system is that the system of the deformable combined robot comprises N AI cell units, wherein the N AI cell units are in closable communication connection, and the arrangement can change data channels among different AI cell units along with environmental change, thereby being beneficial to learning different task contents by AI through different data channel combinations.

Preferably, the AI cell unit comprises an AI processor capable of processing input data to obtain a current cycle processing output, the AI processor being a classifier or a function fitter; an AI predictor for taking at least the current cycle processing data as one of the inputs and outputting prediction data for predicting the next cycle processing output to obtain the next cycle prediction data; and an AI decision maker which takes the processing data of the current period and the prediction data of the next period as input data; the action execution of the AI decision-making device at least comprises an action, namely taking the current period processing data as the input data of some other AI cell unit AI processor, and sending a signal to the distance sensor to position the distance between the connected AI cell unit and the AI cell unit; or the input data received by the AI processor in the current period is changed into the processing data output by some other AI cell unit; the AI cells pointed by the AI decision maker actions are obtained by random search, and the action points are fixed after the initial docking is completed; the decision basis of the AI decision device is that the error between the processing output of the next period and the prediction output of the next period is minimized, through the setting, the AI predictor can learn the data flow mode obtained by the segmented combination in a period of task, so that when the same scene is met, the AI decision device can reappear the channel connection mode in order to minimize the free energy, and a group of AI in the channel connection mode can continue to execute the data flow mode executed before, thereby realizing the effect of decomposing and learning the task under the complex task.

Preferably, the AI cell further comprises a distance sensor capable of sensing the distance between the AI and another AI cell currently in signal connection with the AI cell; the distance data measured by the distance sensor is used as the data input of a motion change decision maker capable of outputting motion to control the movement or peristalsis of the movable layer; the AI processor uses a function fitting device which performs function fitting by using the weight, the weight of the AI processor is used as the input of the shaping layer controller, and the output instruction of the shaping layer controller controls the shape of the shaping supporting layer for the shape data of the shaping supporting layer. The weight of the AI processor is used as the input data of the controller of the shaping layer, so that the shape of the shaping layer can change along with the change of the classification function of the AI processor, and the adaptive change of the robot cell unit on the mechanical function and the shaping function can be realized by matching with the ways of updating the weight of the AI processor or sudden change of the weight of the AI processor and the like.

Preferably, the system further comprises an AI total decision-maker, the AI total decision-maker receives as input at least current cycle processing data in one AI cell unit and prediction data of the next cycle in the current cycle, and the output of the AI total decision-maker acts to select a part of the AI cell units currently connected with the AI total decision-maker, and to disconnect those AI cell units from their data transmission downstream, the arrangement being such that the AI total decision-maker can grasp the flow of the entire data stream according to the task objectives, so that the AI cell units themselves resemble conditioned reflex AI cell units, under the control of the AI total decision-maker, an effect analogous to brain control is obtained, so that the AI has a better alternative capability in performing the task.

As a priority, the decision of the AI total decision maker is based on a task target designated by an operator; or at least calculating interest emotion reward aiming at each connected cell unit, and taking the accumulated interest emotion reward as a decision basis of the AI total decision maker; or all current period processing data output by all AI cell units and all prediction data at the next moment are respectively normalized and then interest emotion rewards are calculated and used as a decision basis of an AI total decision maker, the setting can ensure that the priority can be determined according to own emotion when the AI learns and evolves, and the possibility of evolving towards various directions exists.

Preferably, the action instruction of the AI decision device further includes a mutation action capable of mutating the weight of the AI processor, the mutation action further includes L times of invalid execution times, where L is an integer greater than or equal to 0, the setting enables the AI cell unit to have a mutation mechanism, and by setting the invalid execution times, when the AI can better adapt to the environment to execute tasks, the AI can better execute the data channel switching action due to the effect of the principle of minimizing free energy, and only when the data deviates from the free energy and is minimized, the AI cell unit will lose control and be more likely to have mutation, and if the mutation cannot adapt to the environment well, the mutation will continue in a shorter time until the environment can be adapted finally, and the free energy is minimized.

The invention has the beneficial effects that:

(1) the movable layer is used as a moving means of the cell unit, the shaping supporting layer is used as a supporting framework, and the control chip is used for controlling, so that the cell unit can be moved, combined and changed into a whole with different forms according to needs.

(2) The control chip is arranged according to the principle of minimizing free energy, so that when the classifier transfers data channels between upstream and downstream of different cells, the predictor can learn the data flow of a section of task in the change, and when a cell unit needs to execute a certain section of task, the AI decision-making device drives the cell unit to perform combined change to the task in order to ensure that the classifier data is attached to the prediction period and the free energy is the minimum.

(3) Through the arrangement of a plurality of AI cell units and an AI total decision maker, the relation between the biological brain and the body is simulated, and an AI intelligent agent meeting the requirements of environmental conditions and the task target of the AI total decision maker can be better developed by combining a mutation mechanism.

Drawings

FIG. 1 is a schematic diagram of a cell unit structure of a robot according to the present invention;

FIG. 2 is a schematic diagram of data flow and signal control of the control system of the present invention.

Detailed Description

The invention discloses a machine-deformable combined robot control system, wherein a research team consisting of the university of Egyin Hope and the Kent State university of the United states of the Netherlands has already researched a wriggling material which converts light sensation into displacement, so that on a control chip, only different parts of the wriggling material are irradiated with light, a layer of light isolation material is covered outside the wriggling material in the actual production process, and a layer of solar charging equipment can be arranged on the surface of the light isolation material by internal illumination and is electrically connected with the control chip. The core function of the shaping material is to change the shape, so that an electromechanical structure controlled by a chip, or a short stick driven by a gear to extend and shorten, or two small circular plates can be changed from superposition to separation and also have different supporting frameworks, when the control chip sends different signal data, the structure can be correspondingly changed, so that the cell unit shape is changed, and the cell unit has different motion capabilities by matching with the movable layer.

The first embodiment is as follows:

for simplicity of illustration, three AI cell units and one AI global decision maker are used in this example.

In the actual task execution, in order to receive the external signal, an external signal receiving device needs to be externally connected to convert external data, such as voice, into data that can be processed by the AI cell unit, and an outlet is also needed to output the voice.

In a possible certain time state, an input interface is connected with A, A is connected with B, B is connected with C, C is connected with an output interface, and S controls whether A or B or C is cut off or not.

When the total data characteristics are changed, the free energy of ABC begins to increase, in order to ensure the minimization of the free energy, the decision maker of ABC sends out an instruction to change the data transmission channel, the input interface is connected with A, A is connected with B, A is connected with C, then the output of B and C is fused and then is connected with the output interface, thus the predictor can learn the signal characteristics transmitted under the connection mode, due to the existence of the decision maker, when the data received by the decision maker is changed, the data channel is inevitably changed, thereby ensuring that the predictor can not learn a part of certain possibility, for example, when the ABC is connected in sequence, the data is changed, the free energy is enlarged, at the moment, the predictor can not learn the data transmission content of the changed ABC sequence, but can only learn and predict the data characteristics under the new connection mode after the data channel is changed, thereby ensuring that only one predictor can learn limited content, thereby ensuring that different limited contents are combined to form a whole set of task execution strategy.

In addition, when S is not satisfied with the content of ABC sequence execution, the connection between B and C is cut off, at this moment, C has no upstream data to transmit and stops working, and AB can not transmit data to the output interface, so the influence of the whole agent on the environment is changed, the input data of the input interface is changed, the free energy of AB is increased, and the data channel is further changed, for example, A and B are both connected with the input interface.

In addition, when the characteristics of the input data continuously change, the free energy continuously increases, and at the moment, the decision maker can more easily execute mutation actions, so that weight mutation occurs in ABC after the invalid execution of L, if the mutation can adapt to the environment, the AI tends to be stable, if the mutation can not adapt to the environment, the predictor gradually adapts to the environment in the continuous mutation, better prediction output can be achieved, and the AI also tends to be stable, but the functions of the processor are changed regardless of the type, so that the functions can be more adaptive to the environment.

In addition, the model can be used for unsupervised learning, for example, under an ABC network, a large amount of same voice data is used as input, a predictor of the AI can be gradually adapted to the influence on an AI processor caused by the voice data, at the moment, an output interface of the AI is connected into an input interface of the AI, the self voice of the AI with different characteristics can lean away from the predictor, in order to make the self voice with similar characteristics to training data, the AI can easily mutate towards the voice with similar characteristics to the input data, and simultaneously, the truncation capability of the S in decision ABC is controlled, so that some data can not learn some characteristics without passing through a certain classifier and the predictor, the adverse effect on the whole network when the output data is input in a rotating mode is further eliminated, and after a large amount of data is flushed, the AI can naturally learn to output the data with the original input data characteristics, thereby realizing the unsupervised learning based on the free energy principle.

In the control system of the invention, the processor can adopt a neural network as a function fitting device or a classifier, or can adopt a certain set function to return a value as the input of the next processor, and the predictor can also adopt a neural network, wherein the predictor can adopt a transfomer, a Bayesian neural network and a recurrent neural network as the actual execution tools due to the prediction function of the predictor, and the decision device can adopt reinforcement learning.

In the present invention, the emotional interest mechanism is a method disclosed in the patent with application number 2019112884985, and the difference value of local information between the predicted feature of the next period and the actual environmental feature of the next period is taken as the emotional reward to update the decision maker.

Claims

1. The deformable combined robot is characterized by comprising N cell units, wherein closable communication connection exists among the N cell units, and N is a positive integer.

2. The deformable combined robot is characterized by further comprising a main controller, wherein the main controller is in communication connection with at least one cell unit in the N cell units.

3. A deformable modular robot as claimed in claim 1, wherein said cell unit comprises a control chip, a shaping support layer, and a movable layer connected to said shaping support layer, said control chip being connected in signal communication with said shaping support layer and said movable layer, respectively.

4. A transformable combined robot as claimed in claim 3,

the control chip at least comprises a distance sensor,

a movable layer motion change decision unit,

a shaping change control unit of the shaping supporting layer,

a data processing unit for processing data exchanged with other cell elements,

a data prediction unit for predicting the output data of the data processing unit in the next cycle,

and a behavior unit for switching the transmission channel of the cell unit with other cell units,

the distance sensor is respectively in signal connection with the motion change decision unit and the behavior unit;

the shaping change control unit is in signal connection with the data processing unit,

the data processing unit is in data connection with the data prediction unit;

the data processing unit and the data prediction unit are together in data connection with the behavior unit.

5. The deformable combined type robot control system is characterized by comprising N AI cell units, wherein closable communication connection exists between the N AI cell units.

6. A morphable modular robot control system as claimed in claim 5,

the AI cell unit comprises an AI processor which can process input data to obtain the processing output of the current period, and the AI processor is a classifier or a function fitting device;

an AI predictor for taking at least the current cycle processing data as one of the inputs and outputting prediction data for predicting the next cycle processing output to obtain the next cycle prediction data;

and an AI decision maker which takes the processing data of the current period and the prediction data of the next period as input data;

the action execution of the AI decision-making device at least comprises an action, namely taking the current period processing data as the input data of some other AI cell unit AI processor, and sending a signal to the distance sensor to locate the distance between the connected AI cell unit and the AI cell unit;

or the input data received by the AI processor in the current period is changed into the processing data output by some other AI cell unit;

the AI cell unit pointed by the AI decision maker action is obtained by random search, and the action direction is fixed after the initial docking is completed;

the decision basis of the AI decision maker is that the error between the processing output of the next cycle and the prediction output of the next cycle is minimized.

7. A morphable modular robot control system as claimed in claim 6,

the AI cell also comprises a distance sensor which can sense the distance between other AI cells which are currently connected with the AI cell signal and the AI;

the distance data measured by the distance sensor is used as the data input of a motion change decision maker capable of outputting motion to control the movement or peristalsis of the movable layer;

the AI processor uses a function fitting device which performs function fitting by using the weight, the weight of the AI processor is used as the input of the shaping layer controller, and the output instruction of the shaping layer controller controls the shape of the shaping supporting layer for the shape data of the shaping supporting layer.

8. A morphable modular robotic control system according to claim 6, further comprising an AI total decision maker, said AI total decision maker receiving as input at least current cycle processed data within one AI cell unit and predicted data for a next cycle during a current cycle, said AI total decision maker outputting to select a portion of AI cell units currently connected to the AI total decision maker to disconnect those AI cell units from their data transmission downstream or transmission upstream.

9. A morphable modular robotic control system according to claim 8, wherein the AI total decision maker is configured to base its decision on a task objective specified for the operator;

or at least calculating interest emotion reward aiming at each connected cell unit, and taking the accumulated interest emotion reward as a decision basis of the AI total decision maker;

or normalizing all current period processing data output by all AI cell units and all prediction data at the next moment respectively, and calculating interest emotion reward as a decision basis of the AI total decision maker.

10. The morphable modular robot control system of claim 7 wherein the AI decision maker action instructions further comprise a mutation action that can mutate AI processor weights, the mutation action further comprising L number of invalid executions.