A general biomimetic motion control model based on human brain neuron connection and element shared data architecture and an implementation method thereof

Abstract

The application discloses a general biomimetic motion control model and method based on a human brain neuron connection and element shared data architecture, and belongs to the field of brain-like intelligence and robot motion control. Referring to the human brain motion regulation mechanism, biological units such as skeletal muscle, muscle spindle and Golgi tendon organ are equivalently mapped into servo actuators and matching sensing units; motion and sensing parameters are quantified into basic elements, and each level of element realizes global sharing through indexing; each layer of associated neuron is bound to form a motion primitive and a complex continuous motion sequence. The basic elements support hardware presetting and perception feedback acquired generation; loop synchronous activation enhances synaptic connection efficiency, combined with execution result marking to complete path optimization, to establish a perception-motion correlation memory architecture, and through trial and error iteration, experience solidification and migration are realized. The application is free from dependence on massive labeled data, and the disturbance adaptive capacity is improved, and can be applied to humanoid robots, mechanical arms, exoskeletons and intelligent driving actuators.

Description

Technical Field

[0001] This invention relates to the fields of bionic robot motion control, intelligent automation equipment, and brain-like artificial intelligence. Specifically, it relates to a general bionic motion control model and its implementation method based on a shared data architecture of human brain elements. It is applicable to the underlying motion architecture design, intelligent control model construction, and compliant motion control logic implementation of various multi-joint linkage bionic motion mechanisms such as humanoid robots, multi-degree-of-freedom robotic arms, bionic robotic hands, articulated robots, and wearable assistive exoskeletons. It can also be applied to intelligent driving vehicles to achieve brain-like decision-making and collaborative control of discrete controllable actuators such as steering wheels, accelerators, and brakes. It belongs to the interdisciplinary technical field of brain-like intelligence, robot motion control, and autonomous driving decision-making and execution. Background Technology

[0002] Currently, the intelligent motion control systems of various multi-joint linkage motion devices, such as robots, multi-degree-of-freedom robotic arms, bionic robotic hands, and wearable exoskeletons, are all based on the traditional von Neumann computer architecture and big data-driven artificial intelligence models. The core implementation logic of existing technologies is highly unified: relying on the collection of massive external sample data, offline model training, and global numerical calculations to complete the construction and output of motion strategies, the overall system belongs to a data-driven, computationally intensive machine motion control system.

[0003] This model heavily relies on the pre-accumulation and annotation of large-scale datasets. Whether it's robot motion posture, joint linkage trajectory, or balance adjustment strategy, massive amounts of sample data must be acquired through methods such as manual teaching and sensor data collection. The model then fits and statistically summarizes the patterns of motion parameters based on these samples. Its core logic is that "data scale determines motion performance." Regarding model training mechanisms, existing technologies generally adopt an offline training and online inference mode. All motion logic is solidified into the model parameters during the early training phase. During runtime, only interpolation calculations and trajectory reproduction can be performed, making it impossible to establish adaptive active adjustment logic based on real-time environmental changes.

[0004] At the control execution level, the existing system adopts a global equalization control logic, where all joints are uniformly calculated and synchronously controlled. There is no hierarchical distinction, primary-secondary cooperation, or follow-up adaptive mechanism. Every limb movement relies on the top-level controller to perform real-time large-scale numerical calculations on all degrees of freedom, resulting in extremely high computing power consumption.

[0005] The aforementioned technical approach suffers from three inherent drawbacks: First, its strong data dependency leads to weak generalization ability. For scenarios not covered by training data, the model is prone to trajectory distortion and balance failure, making real-time adaptive adjustment impossible. Second, massive pre-training and global computation result in high development costs and long iteration cycles, hindering the formation of lightweight, universal underlying logic. Third, pure data fitting lacks an intent-driven mechanism; all actions are algorithmic calculations rather than proactive movement logic, resulting in stiff movements, poor coordination, and an inability to achieve natural, fluid human-like movement. The existing system is essentially a replication method based on statistical computation, rather than a hierarchical collaboration and local adaptive control method based on biological neural networks, making it difficult to truly overcome the inherent limitations of mechanical movement.

[0006] Meanwhile, existing academic research on motor neuron mainly focuses on fragmented content such as single neuron firing and local muscle reflexes, which can only explain the principle of single muscle contraction. It has not yet built a complete global theoretical framework for human brain motor regulation from a systems engineering perspective, and cannot guide robots to achieve a complete set of biomimetic motion logic such as hierarchical coordination, local tension adaptation, and prevention of relaxation of basic muscle tension. Summary of the Invention

[0007] I. Purpose of the Invention

[0008] To address the inherent shortcomings of existing motion control systems in robots and multi-joint bionic motion mechanisms, which rely on von Neumann architecture, large-scale pre-training, and globally equalized numerical computation, this invention aims to provide a general bionic motion control model and its implementation method based on a shared data architecture of human brain elements. This model solves the technical problems of existing technologies, such as heavy reliance on massive labeled data, high offline training costs, poor online generalization ability, huge computational overhead, weak motion adaptability, stiff and unnatural movements, and lack of native intention-driven logic. It completely breaks through the technical bottleneck of traditional machine motion control relying on statistical fitting and global computation, constructing a general brain-like bionic motion control system that fully conforms to the native motion regulation mechanism of the human brain. The invention aims to achieve the following objectives:

[0009] 1. Eliminate dependence on large-scale training datasets and improve motion adaptation capabilities in unstructured scenes and under external force disturbances;

[0010] 2. Eliminate global synchronous numerical calculation for all joints to reduce real-time computing power consumption and hardware costs;

[0011] 3. Establish a hierarchical primary-secondary collaborative mechanism driven by consciousness and intention to achieve a human-like, natural, and flexible multi-component linkage effect;

[0012] 4. Relying on biomimetic double-layer closed-loop control, feedback is used to maintain the basic constraint tension, eliminating transmission misalignment and mechanism swaying problems;

[0013] 5. By combining the element-sharing memory framework, we can realize action trial and error learning and solidify proficient actions, thus shortening the equipment action iteration cycle.

[0014] This model is analogous to the mechanism by which the human brain controls the joints of the human limbs. It can be used to control multi-joint bionic mechanisms, or as the "brain" of a machine to manage various discrete and controllable actuators in vehicles and other equipment.

[0015] II. Theoretical Model of the Brain's Motor Nervous System

[0016] Currently, the mechanical hardware processing technology, transmission structure design, and joint degree-of-freedom layout of various bionic motion mechanisms have reached a relatively high level of maturity, and the physical difference between the hardware and human limbs is relatively limited. The core bottleneck restricting the widespread application and generalization of such devices lies not in the mechanical hardware itself, but in the fundamental misalignment between the operating logic of the motion control system and the original motion regulation mechanism of the human brain.

[0017] This invention is based on a systematic study of the structure of the human brain's biological neural network, the mechanism of motor nerve regulation, and the laws of human motor learning. It imitates the original working mechanism of human brain limb movement and combines the engineering features of a multi-joint mechanical control system to achieve the implementation of the technology.

[0018] The theoretical model of the human brain motor neural network upon which this invention is based is a systematic and original theory formed by the inventor after long-term research on the mechanism of consciousness and the human neural regulation system, integrating systems theory and the principles of communication information transmission, and summarizing and condensing existing research results in motor neuronology. The element-sharing data processing infrastructure in this theoretical model has been fully disclosed in the inventor's Chinese invention patent (application publication number: CN122133731A) published on June 2, 2026. All content in that document regarding quantified neurons, element storage indexes, signal aggregation, and memory retrieval is incorporated herein by reference. This invention is a specific application scheme further extended to the field of artificial motor control within the framework of the aforementioned theoretical model. The theoretical model itself does not constitute a disclosure of the technical solution of this application; this application protects the specific implementation method, control logic, and system architecture based on this theory.

[0019] To clearly illustrate the subsequent technical solutions, the following is an overview of the core content within this theoretical framework that is strongly related to motion control.

[0020] (I) Basic Structure and Components of the Human Musculoskeletal System

[0021] The complete human limb movement system is composed of bones, joints, skeletal muscles, tendons, and two types of proprioceptors. Bones act as rigid support structures, joints provide multi-degree-of-freedom articulation, skeletal muscles are the sole actuators, generating traction force through contraction and relaxation, and tendons bind skeletal muscles to the skeleton to transmit tension. Skeletal muscles themselves lack sensory capabilities; their perception of movement relies on two specialized sensory units: muscle spindles (sensing muscle extension length and rate) and the Golgi tendon organ (sensing muscle contraction tension and external load). Joints and bones are conventional mechanical structures and will not be discussed in detail in this paper. The subsequent focus will be on the dynamic characteristics of skeletal muscles and the structure and function of muscle spindles and the Golgi tendon organ.

[0022] 1. Structure, function, and dynamic characteristics of skeletal muscle

[0023] Skeletal muscle is the only active power execution unit in the human musculoskeletal system. It is composed of a large number of extrafusal muscle fibers that aggregate to form the muscle belly, and its two ends are anchored to the bones through tendons, forming a rigid and interconnected mechanical kinematic chain between muscles, bones, and joints. Skeletal muscle is a purely power execution structure, only responsible for the functions of contraction under force and output of tension, and does not contain autonomous sensing and judgment modules.

[0024] Skeletal muscles primarily comprise three basic functions: First, joint drive function, which generates traction force through the active contraction and relaxation of muscle fibers to drive joint rotation; second, velocity regulation function, where the contraction and relaxation rates directly determine the speed and rhythm of joint movement; and third, postural support function, which maintains posture by sustaining continuous micro-contraction tension to counteract body weight and load. Skeletal muscles exhibit typical unidirectional power output, imperceptibility, and lack of self-closed-loop regulation. All their precise positioning, velocity control, and load adaptation capabilities rely on the sensory feedback from muscle spindles and the Golgi tendon organ.

[0025] 2. Structure, function, and manipulation characteristics of muscle spindles

[0026] Muscle spindles are the core proprioceptors in human skeletal muscles that are responsible for motion perception. They are embedded in the skeletal muscle group and are arranged in parallel with the main muscle fibers. They form a rigid synchronous relationship with the muscle, and their deformation and deformation speed are completely consistent with the actual movement state of the joint, thereby achieving accurate perception of the limb movement state.

[0027] Muscle spindles possess two core sensory capabilities: First, length extension sensing (position sensing), which generates nerve impulses of corresponding intensity based on the total muscle stretch deformation, characterizing the current spatial position of the joint and providing a positional reference for the musculoskeletal system; second, stretch rate sensing (velocity sensing), which identifies the changes in the speed of muscle stretching and contraction in real time, converting the deformation rate into nerve firing frequency to achieve dynamic control of movement rhythm. In summary, muscle spindles fully undertake the functions of position detection, velocity detection, and movement process regulation, serving as the original biomimetic prototype for the closed-loop position and velocity mechanisms of biological movement.

[0028] 3. Structure, function, and manipulation characteristics of the Golgi tendon organ

[0029] The Golgi tendon organ is a proprioceptor specifically responsible for mechanical sensing and load regulation. It is located in the transmission pathway connecting muscles and tendons, and is in series with the main muscle fibers. All muscle contraction tension and external loads must pass through its sensing structure, enabling it to independently and accurately collect the tension in the transmission link.

[0030] The Golgi tendon organ focuses on two core capabilities in the mechanical dimension: first, muscle contraction tension sensing, which detects the output force of active muscle contraction in real time; second, external load and gravitational disturbance sensing, which sensitively captures changes in limb weight and external load. In terms of maneuverability, the Golgi tendon organ constructs a load-adaptive force balance closed loop: during dynamic movement, it monitors tendon tension changes in real time, constraining muscle output force to avoid overload; during steady-state support, it continuously provides feedback on load tension, fine-tuning muscle contraction intensity to maintain a dynamic balance between output tension and external resistance, achieving high-precision steady-state load holding. In summary, the Golgi tendon organ fully undertakes the functions of tension detection, load recognition, and mechanical adaptive regulation, serving as a native biomimetic prototype of a biological motion torque closed loop.

[0031] (II) Basic Motion Units in the Human Motion System

[0032] The motor unit is the smallest controllable unit of skeletal muscle, consisting of a single α motor neuron, corresponding nerve fibers, and innervating muscle fibers. Muscle fibers within the same motor unit contract and relax synchronously, while the control pathways of different units are independent of each other, allowing the central nervous system to selectively group and regulate units.

[0033] The motor unit possesses four core characteristics: synchronous linkage within the unit, independent controllability between units, a physiological upper limit to the output force of a single unit, and dual sensory constraints from the built-in muscle spindle and Golgi tendon organs. All skeletal muscles in the human body are composed of a large number of motor units, and the nervous system completes graded regulation of force by activating the number of units, pulse frequency, and activation sequence.

[0034] (III) Bionic Model for Human Motion Control

[0035] 1. Classification of motion states

[0036] Based on human limb behavior, movement is divided into four categories. This invention only models the first three categories of brain-dominated movement; spinal cord stress reflexes are not within the scope of this study.

[0037] (1) Active controlled movement: The brain issues directional control commands, designating the execution unit to complete the purposeful action according to the target trajectory, speed, and force, which is the core operation movement mode;

[0038] (2) Linked following motion: The central unit only issues active commands to the core execution unit, while the other units are passively displaced by mechanical coupling. The corresponding execution unit adapts and adjusts its output based on tension feedback to maintain smooth following and avoid jamming and stress overload.

[0039] (3) Postural steady state maintenance: After active movement is completed, the execution unit switches to the local tension closed loop and relies on feedback to dynamically balance the external load; only the tension closed loop is effective and does not automatically correct small deviations. Large deviations require the brain to reissue active commands to reset.

[0040] (4) Stress-protective reflex: triggered independently by the underlying layer, without the need for brain decision-making, belongs to the underlying safety mechanism and will not be elaborated.

[0041] 2. Theoretical Model of Kinematic Unit

[0042] A single motion unit constructs a layered, dual-closed-loop control architecture: the inner layer is a tension closed loop formed by the Golgi tendon organs, and the outer layer is a position and velocity closed loop formed by muscle spindles; the outer loop outputs the main control signal, and the signals from both the inner and outer loops are aggregated into a unified signal integration unit, operating in three modes, such as... Figure 1 As shown:

[0043] (a) Increment of outer ring main control signal: The inner and outer rings work together in a dual closed-loop operation to correct the execution status in real time and synchronously complete tension overload protection;

[0044] (b) The outer loop main control signal is constant: the dual closed loop continuous negative feedback regulation cancels out external disturbances and stabilizes the attitude;

[0045] (c) The outer ring main control signal is zeroed: the outer ring link is locked and stops working, and only the inner tension closed loop operates independently; relying on real-time tension feedback to adaptively adjust the output, the mechanism can be smoothly driven by external force, while maintaining weak tension to avoid complete relaxation and shaking.

[0046] After the outer ring is closed, a three-layer safety mechanism works together to protect the system: the inner layer provides tension closed-loop constraint stress; the mechanical limit mechanism completes the physical safety net; and the internal sensing continuously feeds back the status to the decision-making level, issuing active braking commands from the central control unit in case of an anomaly.

[0047] The entire dual-loop architecture is an innate neural structure; only the upper limit of unit muscle strength can be fine-tuned through training, while the underlying closed-loop structure cannot be added later. This model decomposes the signal integration logic of biological single neurons into a modular expression of two layers of controllable signal sources, which is merely an engineering simplification method and does not violate the original biological regulatory logic.

[0048] 3. Definition of muscle functional blocks

[0049] A single anatomical muscle contains numerous motor units, and directly using the entire muscle as a control unit leads to insufficient control precision. This invention, without dissecting the boundaries of the anatomical muscle, logically groups the internal motor units according to the movement task to form muscle functional blocks. A single muscle can be divided into one or more functional blocks, with the central nervous system issuing independent master control signals to each block. Fine-grained multi-condition control is achieved through a dual closed-loop system, simplifying the scheduling logic of the whole-body execution units. Its structure is shown below. Figure 2 .

[0050] 4. Overall Regulation Model of the Motor Nervous System

[0051] A complete movement can be broken down into continuous movement slices along the time axis. Within a single slice, the amplitude of the main control signal for each functional block remains constant, and signals from adjacent slices transition smoothly to avoid jitter. Slice parameters are stored in a shared memory database, allowing skilled movements to be directly retrieved and reproduced using slice sequences. The overall regulatory model of the motor nervous system constructed based on this logic is as follows: Figure 3 As shown.

[0052] The system is divided into six independent functional areas, each performing its own function to complete the entire process of perception, decision-making, storage, and command output:

[0053] ① Signal aggregation area: gathers internal and external sensory signals, connects with the occupancy and quantification neurons of the prior element sharing architecture, and synchronously receives historical information transmitted back from the memory area; input sensory signals include visual signals, auditory signals, tactile signals, internal sensory signals, and body balance signals;

[0054] ② Sensory generation zone: distinguishes pulse signals from different pathways and generates various subjective perceptions;

[0055] ③ Memory area: Based on the prior element sharing architecture, the storage and retrieval of motion scenes and motion sensations are completed;

[0056] ④ Consciousness and decision-making center: Recognizes perceived information, controls memory reading and writing, and outputs high-level motor intentions;

[0057] ⑤ Motion instruction parsing area: Deconstructs abstract motion intentions, filters functional blocks involved in the action, and generates hierarchical main control signals;

[0058] ⑥ Action Execution Area: Contains all functional blocks, receives main control commands and underlying perception feedback, and completes layered closed-loop control.

[0059] Two independent feedback chains are distinguished: the functional block has a built-in dual closed loop for real-time correction feedback at the bottom layer; the global perception feedback sent to the signal collection area does not participate in the bottom layer adjustment, but is only used to establish scene-action memory association, providing a basis for high-level decision-making and strategy adjustment.

[0060] 5. Typical characteristics of this biological movement model

[0061] (1) Layered feedback isolation: The bottom layer double closed loop ensures execution stability, while the top layer somatosensory feedback is used for motor learning and memory. The two functions are completely separated.

[0062] (2) Element-based motor learning: New movements are executed through trial and error → feedback correction → path selection → memory solidification process, and the optimal movement path is continuously reinforced;

[0063] (3) Motion-triggered proficient actions: Proficient actions are bound to exclusive motion-sensing elements. During operation, there is no need for central frame-by-frame control. Only intermittent supervision and intervention are required, reducing the consumption of high-level computing power.

[0064] (4) Equivalent mapping of biological structure engineering: There is no need to replicate the anatomical morphology of neurons, only the functional logic is abstracted, and motor units and receptors can directly correspond to the controllable execution components and sensing modules of various devices.

[0065] III. Technical Solution

[0066] This invention is based on a previously disclosed human brain element-sharing data processing architecture, and engineered the hierarchical regulation of human motor nerves, in vivo perception closed loop, and action learning mechanism. Combining computer hardware characteristics, it transforms biological neural connections into a computable data association structure, establishing an equivalent mapping between biological motion units and controllable actuators in equipment, thus building a biomimetic hierarchical motion control system. This architecture is universal and can be used for multi-joint mechanisms in robots, as well as various discrete controllable actuators in vehicles and other equipment.

[0067] The underlying data processing infrastructure for element sharing has been fully disclosed in the applicant's published Chinese invention patent (application publication number: CN122133731A). All content in that document concerning quantized neurons, element storage indexes, signal aggregation, and memory retrieval is incorporated herein by reference. Based on the aforementioned architecture, this invention further proposes a proprietary biomimetic mapping and control process for the field of motion control.

[0068] Considering that computer hardware lacks the real-time dynamic plasticity of human brain neurons, this invention combines the characteristics of computer hardware architecture with engineering control requirements. It unifies the biophysical connection between human brain neurons and motor nerve units into a data structure and element association that can be described and computed by computers. At the same time, it establishes a precise equivalent mapping between biological motor units and controllable hardware execution components. It equivalently transplants the proprioception and regulation mechanisms of human muscle units, muscle functional blocks, muscle spindle receptors, and Golgi tendon organs into a multi-dimensional sensing and acquisition and dual-layer closed-loop regulation architecture for the position, rotation speed, and torque / force of the execution unit. Relying on element sharing, simplified neural conduction, and trial-and-error iteration mechanisms, it realizes the complete engineering implementation of biomimetic human brain cognitive logic and biomimetic motor nerve regulation mechanisms in the control system.

[0069] (I) Equivalent mapping relationship between biological musculoskeletal system and hardware

[0070] 1. Equivalent correspondence between skeletal muscle and servo actuator

[0071] In biological systems, skeletal muscle is the sole active power actuator of the human musculoskeletal system, undertaking three basic functions: joint drive, speed regulation, and posture support. It possesses the characteristics of unidirectional power output, imperceptibility, and lack of self-closed-loop regulation. In engineering systems, servo motors, vehicle steering wheels, brakes, accelerators, and other servo actuators, as power output components, also perform power output functions. Their output force / stroke can be controlled and adjusted, but they lack autonomous sensing and error correction capabilities. Therefore, skeletal muscle is equivalently mapped to various servo actuators; the traction force generated by muscle contraction and relaxation is equivalent to the output driving force of the actuator; the muscle contraction rate is equivalent to the action rate of the actuator; and the sustained muscle contraction to maintain posture is equivalent to the locking and force-holding output states of the actuator.

[0072] 2. Equivalent correspondence between muscle spindles and position detection units (including differential calculation)

[0073] In biological systems, muscle spindles are embedded within skeletal muscles, forming a parallel structure with the main muscle fibers. They can simultaneously sense muscle extension length (corresponding to joint spatial position) and extension rate (corresponding to joint movement velocity), providing positional reference and velocity feedback for the musculoskeletal system. In engineering systems, encoders, angle sensors, and stroke sensors collect the position of the actuators in real time, and obtain motion rate information through differential calculation of position data. Therefore, muscle spindles are equivalently mapped to position sensors and their associated differential calculation units; the muscle spindle's extension length sensing function is equivalent to position readings, and its extension rate sensing function is equivalent to the rate information obtained from differential position data.

[0074] 3. Equivalent correspondence between Golgi tendon organs and force / torque sensors

[0075] In biological systems, the Golgi tendon organ is located in the transmission pathway connecting muscles and tendons, forming a series structure with the main muscle fibers. All muscle contraction tension and external loads must pass through its sensing structure, specializing in the real-time detection of muscle contraction tension and external load disturbances. In engineering systems, force sensors and torque sensors are arranged in series in the force transmission link to collect changes in the force applied in the transmission link. Therefore, the Golgi tendon organ is equivalent to a force / torque sensor, tension sensing is equivalent to the force measured by the sensor, external load disturbance sensing is equivalent to the external disturbance signal captured by the sensor, and load adaptive regulation is equivalent to closed-loop regulation based on force feedback.

[0076] 4. Equivalent correspondence between motion unit and single independent drive execution unit

[0077] In biological systems, the motor unit is the smallest controllable functional unit that can be independently scheduled and controlled by the nervous system. It possesses four inherent characteristics: overall synchronous linkage of units, independent controllability between units, graded adjustment of force, and built-in dual-sensor cooperative dynamic constraints. In engineering systems, a single independently driven actuator and its associated sensors constitute a complete, independent, and controllable execution channel. It can independently receive commands from the upper level to complete actions, and the control links between channels are isolated from each other. Therefore, a single motor unit is equivalently mapped to a set of independently driven actuators and their associated sensing components.

[0078] 5. Further equivalence mapping between muscle functional blocks and independent execution channels

[0079] In biological systems, a muscle functional block is an internal control subunit formed by logically grouping the motor units within a single anatomical muscle without disrupting its physical boundaries. In engineering, each independent execution channel can be equivalent to a muscle functional block. The complex movements of the entire machine are achieved by combining multiple independent execution channels in a parallel and coordinated manner.

[0080] In summary, the position, speed, and force / torque sensing parameters of the execution unit completely and equivalently replace the sensing functions of biological muscle spindles and Golgi tendons, constituting a biomimetic somatosensory data source; each independent execution channel simultaneously assumes the dual equivalent roles of a motion unit and a muscle functional block.

[0081] (II) Rules for Element-based Processing of Motion Control Parameters

[0082] Based on the shared data processing architecture of preceding human brain elements, this invention establishes unified parameter elementization and quantification hierarchical processing rules adapted to engineering control, targeting the signal transmission, feedback regulation and command output characteristics of the motor nervous system.

[0083] In the human motor nervous system, all information is transmitted in the form of pulse signals, without continuous floating-point numerical operations. This invention replicates this biological signal mechanism, abandoning the traditional control method that relies on complex floating-point operations and continuous numerical adjustments. It maps all input, feedback, and output parameters of the control system to quantized neurons, completing signal expression, memory storage, and command driving through discrete levels. All quantization levels and grading intervals are not generated by algorithm fitting but are entirely predetermined by the hardware circuit range, sensor sampling range, output drive range, and system control precision requirements. Each type of signal corresponds to an independent and fixed quantization level system. The specific processing principles are as follows:

[0084] 1. Quantitative and element-based processing of externally perceived information

[0085] External environmental perception signals, such as visual, auditory, and tactile signals, are transmitted to the central nervous system in the form of neural impulses. Based on the sensor hardware sampling range, signal voltage extreme values, and sampling accuracy requirements, a fixed number of quantization levels are defined. Continuous sensory physical signals are transformed into discrete, indexable, and memorizable standardized sensory elements, which are stored in a global element memory bank, providing a unified element data source for subsequent motion decision-making and action matching.

[0086] 2. Quantitative and element-based processing of ontological feedback information

[0087] (1) Force / torque feedback quantization rules: Based on the detection range of force and torque sensor hardware, the amplification coefficient of feedback circuit, the minimum identification threshold and the maximum load extreme value, the continuously changing physical quantity of force is divided into multiple discrete quantization levels according to the system control precision requirements. The current load state is represented by the quantization neuron level, forming the force elements.

[0088] (2) Position and rate feedback quantization rules: Based on the position sensor resolution, actuator stroke extreme value, rate upper and lower limits and motion smoothness accuracy requirements, the position offset and action rate are quantized and graded at fixed levels to form attitude elements.

[0089] 3. Quantization and element-based processing of motion master control output commands

[0090] The high-level decision-making center generates intentional commands regarding force intensity and amplitude based on the motion target, and pre-divides fixed drive quantization levels according to the range of the driving hardware, speed adjustment range, and output smoothness level. Different control intentions for different motion intensities and amplitudes activate different levels of quantization neurons, and the final output drive level is determined by the level of the quantization neurons, realizing a complete biomimetic output chain from "high-level intention - quantization expression - hierarchical drive of the actuator".

[0091] 4. The general principle of unified elements

[0092] All sensing parameters, feedback parameters, and output control parameters of the control system of this invention follow the same set of biomimetic rules:

[0093] (1) The signal transmission layer continues the characteristics of biological pulse signals, without continuous numerical calculations;

[0094] (2) The signal expression level relies entirely on quantized neurons for discrete-level expression;

[0095] (3) All levels and grades at the quantitative basis level are determined by hardware physical parameters, rather than generated by software algorithms;

[0096] (4) All quantified motion elements are uniformly incorporated into the element sharing architecture at the data storage level, with the bottom layer storing entities and the upper layer indexing, to achieve complete unification of motion control and bionic brain cognitive architecture. Attached Figure Description

[0097] Figure 1 This is a schematic diagram of a motion unit model according to Embodiment 1 of the present invention, including... Figure 1 (a) Figure 1 (b) Figure 1 (a) is a detailed block diagram of the motor unit model. The main control signal C is input to the controllable signal source 2, and the output signal C2 of the controllable signal source 2 is sent to the controllable signal source 1. The output signal C1 of the controllable signal source 1 is sent to the muscle motor unit, and the muscle motor unit outputs a motor output. The motor output leads to two feedback paths. One path is sent to the Golgi tendon organ, and the Golgi tendon organ outputs a feedback signal F1 to the controllable signal source 1. The other path is sent to the muscle spindle, and the muscle spindle outputs a feedback signal F2 to the controllable signal source 2. Figure 1 (b) is Figure 1 (a) is a simplified block diagram. The main control signal C is directly input to the motion unit, and the motion unit outputs motion output.

[0098] Figure 2 This is a schematic diagram of the muscle functional block according to Embodiment 1 of the present invention. The muscle functional block is completely surrounded by a large dashed frame. Inside the dashed frame, motor unit 1 and motor unit 2 are arranged sequentially from top to bottom, and motor unit N is arranged after the vertical dashed line is omitted. A circular node representing a neuron is provided on the left side of the dashed frame, and the main control signal C points to this circular node through an arrow. Multiple arrowed lines are led out from this circular node, pointing to motor unit 1, motor unit 2 and then to motor unit N respectively. A summation circular node with a plus sign inside is provided on the right side of the dashed frame. The output signals of motor unit 1, motor unit 2 and motor unit N are respectively connected to this summation circular node to complete the synthesis of the motor output of each motor unit. Arrows are led outward from the summation circular node to output the motor output.

[0099] Figure 3This is a schematic diagram of the motor nervous system regulation principle according to Embodiment 1 of the present invention. A signal gathering area box is located in the upper left corner of the diagram. Above the signal gathering area are five parallel input arrows, representing visual, auditory, tactile, internal sensory, and balance signals, respectively. Below the signal gathering area are parallel boxes for a sensory generation area and a memory area, with arrows extending from the signal gathering area pointing to the sensory generation area and the memory area, respectively. An arrow from the sensory generation area points to the circular node of the consciousness decision-making center, and an arrow from the consciousness decision-making center points to the memory area. To the right of the consciousness decision-making center is a movement command parsing box, connected to the signal gathering area by a rightward arrow. To the right of the above structure is a large dashed box serving as the action execution area. Inside the action execution area, muscle function blocks 1 and 2 are arranged sequentially from top to bottom, and muscle function block M is arranged after the vertical dashed line is omitted. Arrows connecting the signal gathering area and the movement command parsing box point to the dashed box of the action execution area.

[0100] Figure 4 This is a schematic diagram of the control mode of a single motion unit according to Embodiment 1 of the present invention. Figure 4 The system is divided into a basic element layer W0 and an associated layer W1. The W0 layer contains three sets of basic elements enclosed by dashed boxes: main control signal C, motor operating mode M, and motion unit terminal block T. The main control signal C is exemplarily divided into four quantization levels: U0, U1, U2, and U3. The motor operating mode M is exemplarily divided into four operating modes: U4, U5, U6, and U7. The motion unit terminal block T is exemplarily configured with six terminals: U8, U9, U10, U11, U12, and U13. Layer 1 has 7 associated neurons, namely L10, L11, L12, L13, L14, L15, and L16. The connection relationships between each associated neuron and the basic elements of layer W0 are as follows: L10 connects to U0, U5, and U8; L11 connects to U0, U6, and U9; L12 connects to U1, U5, and U10; L13 connects to U2, U6, and U11; L14 connects to U3, U7, and U12; L15 connects to U3, U6, and U13; and L16 connects to U3, U7, and U13.

[0101] Figure 5 This is a schematic diagram of the passive following mode of a single motion unit in Embodiment 2 of the present invention. Figure 5The system is divided into a basic element layer W0 and an associated layer W1. The W0 layer contains three groups of basic elements enclosed by dashed boxes: torque feedback group FT, motor operating mode M, and motion unit terminal block T. The torque feedback group FT is exemplarily divided into four quantization levels: U14, U15, U16, and U17. The motor operating mode M is U4, U5, U6, and U7. The motion unit terminal block T is U8, U9, U10, U11, U12, and U13. The W1 layer has seven associated neural networks. The neurons are L17, L18, L19, L20, L21, L22, and L23. The connection relationships between each associated neuron and the basic elements of layer W0 are as follows: L17 connects to U14, U4, and U8; L18 connects to U15, U4, and U9; L19 connects to U14, U4, and U10; L20 connects to U16, U4, and U11; L21 connects to U17, U4, and U12; L22 connects to U16, U4, and U13; and L23 connects to U17, U4, and U13.

[0102] Figure 6 This is a schematic diagram illustrating the formation of continuous action in Embodiment 3 of the present invention. Figure 6 It is divided into three interconnected layers: W1, W2, and W3. Neurons in layer W1 include L10, L11, L12, L13, L18, L19, L25, L35, and L40; neurons in layer W2 include L20, L21, L22, L23, L24, L25, L26, L27, and L28; neurons in layer W3 include L30, L31, L32, L33, L34, L35, L36, L37, and L38. The connections between layers W1 and W2 are as follows: L20 connects to L10 and L11; L21 connects to L12, L13, and L18; L22 connects to L10 and L19; L23 connects to... Connect L11 and L25; L24 connects to L12 and L19; L25 connects to L13 and L18; L26 connects to L18 and L35; L27 connects to L25 and L35; L28 connects to L19 and L40; Connection relationship between W2 and W3 layers: L30 connects to L20 and L21; L31 connects to L22 and L23; L32 connects to L20 and L24; L33 connects to L22 and L25; L34 connects to L24 and L26; L35 connects to L23 and L25; L36 connects to L27 and L28; L37 connects to L25, L26, and L28; L38 connects to L21 and L27.

[0103] Figure 7 This is a schematic diagram illustrating the association between perceptual feedback and motor loop memory in Embodiment 4 of the present invention. Figure 7The system is divided into three layers: WX, WY, and WZ. The WX layer contains two modules, A and B, enclosed by dashed boxes. Module A contains A0-A11, and module B contains B0-B4. The WY layer contains twelve neurons, LY0-LY11. The WZ layer contains seven neurons, LZ0-LZ6. The connections between the WX and WY layers are as follows: LY0 connects to A0 and A5; LY1 connects to A1 and A6; LY2 connects to A2 and A7; LY3 connects to A3 and A8; LY4 connects to A4 and A9; LY5 connects to A10 and B0; LY6 connects to A10 and B0. Connect A10 and B1; LY7 connects to A10 and B2; LY8 connects to A10 and B3; LY9 connects to A11 and B2; LY10 connects to A11 and B0; LY11 connects to A11 and B4; Connection relationship between WY layer and WZ layer: LZ0 connects to LY0 and LY5; LZ1 connects to LY1 and LY6; LZ2 connects to LY2 and LY7; LZ3 connects to LY3 and LY8; LZ4 connects to LY4 and LY11; LZ5 connects to LY2 and LY9; LZ6 connects to LY0 and LY10. Detailed Implementation

[0104] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The embodiments focus on how to determine the basic motion elements according to the processing object, how to divide the sub-elements at each level according to the preset hierarchical structure, and what structure and method are used to realize the association and sharing between motion elements and sub-elements at each level; at the same time, it demonstrates the mechanism of simulating neuronal plasticity in artificial systems, establishing connections between elements in real time according to actual needs, and realizing dynamic adjustment of neuronal synaptic connection efficiency.

[0105] The following examples use a general robot mechanical structure and drive control system as the descriptive paradigm. The drive control methods and sensing feedback forms involved are set according to common knowledge in the field of electromechanical engineering, without limiting the specific model, power, or specifications of the motor. For example, the motor regulates its operating state by pulse frequency and adjusts its output force by input voltage. The speed feedback equivalent to the muscle spindle and the torque feedback equivalent to the Golgi tendon organ use feedback voltage as the signal carrier. The hierarchical quantization method of feedback voltage and motor drive voltage are all general control logic and are not bound to any fixed hardware values.

[0106] It should be noted that the control logic described in this invention pertains to behavioral regulation logic at the level of brain consciousness decision-making and does not involve the underlying drive implementation within the actuator. This invention only sets behavioral requirements for the actuator: the actuator should be able to complete motion output according to the speed, torque, and other indicators corresponding to the control commands issued from the upper layer. How the actuator internally achieves this response effect is beyond the scope of this invention. If brushless DC motors or brushed DC motors, which do not inherently possess closed-loop feedback, are selected, then additional sensing feedback units equivalent to muscle spindles and Golgi tendons need to be added at the hardware level. This relies on the underlying hardware closed loop to ensure that the corresponding motion state can be maintained under a given voltage, such as automatically adjusting the output power to maintain stable operating speed when the load changes. However, the servo motors commonly used in humanoid robots already integrate position, speed, and torque closed-loop control. This underlying closed-loop function is a built-in hardware function, and this invention will not discuss its internal implementation.

[0107] The control model described in this invention is constructed based on the hierarchical structure and information processing mechanism of the brain's neural network; therefore, biomimetic terms such as "consciousness," "sensation," and "perception" are used in the description. It is important to note that these terms have specific engineering meanings in this invention and do not imply that the robot possesses biological subjective consciousness or sensory abilities. In the engineering implementation of this invention, "consciousness" corresponds to the decision-making logic and control flow in the program. For example, "establishing a connection between two neurons" is accomplished by the brain in biological systems, but in this invention, it is defined by the program code—the program explicitly specifies which units should establish connections, what kind of index should be created, and what index format should be used. All operations at the "consciousness level" ultimately manifest as judgment, selection, and execution logic in the program.

[0108] "Sensation" and "perception" correspond to the various types of information fed back during system operation. Machines themselves do not possess biological sensory abilities, but they can acquire external and internal state data through sensors—including joint angles, torque, speed, acceleration, visual image features, tactile pressure values, etc. In this invention, this data is stored in a unified encoded form within the basic elements of the WO layer, serving as the basis for system decision-making and learning. The so-called "sensory memory," in engineering implementation, is the storage and retrieval of the aforementioned encoded feedback information.

[0109] Therefore, the terms "consciousness," "sensation," and "perception" used in this specification should be understood as functional analogies to the working mechanisms of biological motor systems, rather than descriptions of a robot system possessing subjective consciousness or sensory abilities. In actual engineering implementation, all "consciousness" activities correspond to program logic, and all "sensation" corresponds to coded sensor feedback data.

[0110] To clearly explain the core mechanisms of the quantization neuron mapping rules, the motion dual-loop control architecture, and the element-based processing of control parameters, the example demonstrates the preset voltage range and signal range, and divides the quantization levels according to the control precision. This range and level division is for illustrative purposes only; in actual engineering applications, those skilled in the art can set the upper and lower limits of the signal themselves based on the inherent hardware attributes of the drive motor, such as its voltage range, power range, speed adjustment range, and torque detection range, and redefine the quantization levels according to the actual control precision requirements.

[0111] Even in more complex actual working conditions, with more sensing parameters and constraints involved in closed-loop control, and diverse robot body structures, as long as the core architecture of this invention is followed—external sensing signals, body feedback signals, and main control output commands are uniformly encoded through quantized neurons, and the quantization level is determined based on the hardware range and control precision—this biomimetic motion control can be achieved. Furthermore, in addition to its universality at the hardware parameter level, this invention also possesses high flexibility in selecting upper-level internal control strategies.

[0112] For example, while servo motors commonly used in humanoid robots integrate position and velocity dual closed loops within their hardware, enabling stable movement under given commands, they lack the biomimetic servo control capabilities of human muscles. When the upper-level master control signal is withdrawn, the servo either enters a zero-torque mode, exhibiting near-complete freedom, or locks itself in its current position, unable to autonomously generate the partially relaxed, dynamically maintained, slightly tense servo effect of human muscles under external force. This invention, however, at the brain-like upper-level control level, dynamically issues small-amplitude following torque commands based on the tension signals received from the passive motion state, thereby simulating the body's basic muscle tension and preventing excessive joint relaxation and swaying.

[0113] Taking the dual-closed-loop control logic of the motion unit of this invention as an example, the basic paradigm of this invention is as follows: when the upper-level master control signal disappears, the outer closed loop responsible for tension speed and position control stops working, and only the inner loop adjusts instantaneously according to the real-time tension, so that the actuator moves smoothly under the influence of external force, avoiding rigid resistance. Based on this, those skilled in the art can flexibly select other control schemes according to the operating conditions such as equipment operation requirements, mechanical protection requirements, and motion smoothness optimization goals, without changing the core principles such as the overall architecture of the dual closed loop and the element processing mechanism. For example, an additional restricted working logic for the outer loop can be set after the master control signal disappears, so as to moderately constrain the passive tension speed through the outer loop, reduce mechanical impact, and extend the structural life; such differentiated control schemes can all be extended and optimized within the overall framework of this invention, reflecting the openness and customizability of the control system architecture.

[0114] I. Example 1: Driving Implementation of a Single Motion Unit

[0115] Figure 4 The diagram shown is a neural network structure diagram of this embodiment, providing a brain-like neural hierarchical drive control method for a single motion unit applicable to a robot biomimetic motion system. The overall architecture is strictly divided into a two-layer structure: the W0 basic element layer and the W1 associated neuron layer. All control parameters, operating modes, and output ports are stored in the bottom layer in the form of standardized elements, and then a "CMT" control mode triplet is formed through the W1 associated neurons for a single motion unit.

[0116] (I) Construction of W0 Basic Element Layer

[0117] Layer W0 is a globally shared basic element layer. The parts relevant to this embodiment are divided into three independent element sets: main control signal group C, motor operating mode group M, and motion unit terminal output group T. The elements in each group are as follows:

[0118] 1. Main control signal group C (quantization voltage element)

[0119] Figure 4 The signal is symbolically divided into four quantization levels, denoted as U0, U1, U2, and U3. In practical engineering, the main control signal C is represented as an analog voltage command applied to the motor drive. For bidirectional robot joint motors, the voltage range is exemplarily set to -10V to +10V, where positive and negative polarities correspond to two opposite rotational directions of the motor rotor. It should be noted that "positive" and "reverse" are not limitations on the absolute direction of motor rotation, but rather a set of relative definitions—that is, a first rotational direction and a second rotational direction, which are opposite to each other. Those skilled in the art can determine which direction is positive and which is negative based on the specific motor model, installation method, and reference coordinate system of the motion system; the scope of protection of this invention is not affected by this determination method.

[0120] Depending on the actual requirements of the application scenario or control precision, the quantization value of the main control signal C can be generated using any of the following methods or a combination thereof:

[0121] (1) Pre-set method

[0122] That is, during the system initialization phase, the quantization level and numerical range of the main control signal C are pre-set based on the motor's rated parameters and typical operating conditions. The quantization division can adopt any of the following strategies:

[0123] (a) Uniform division: The entire voltage range is uniformly divided into multiple quantization levels at fixed intervals (e.g., 0.5V), with each level connected end to end and distributed at equal intervals;

[0124] (b) Nonlinear division: Based on the operating characteristic curve of the motor or the specific accuracy requirements of the motion system, different quantization intervals are used in different voltage ranges. For example, in the low-speed range or the range of sudden torque changes where the motor is sensitive to voltage changes, a smaller quantization interval is set to improve control resolution; in the high-speed range or the nonlinear range where the motor response tends to saturate, a larger quantization interval is set to save control resources. In addition, if a specific motion range (such as the fine-tuning section where the end approaches the target position) requires fine control, a smaller quantization interval can also be set separately in that range.

[0125] The aforementioned preset method corresponds to the genetically determined innate reflexes and instinctive response mechanisms in the biological nervous system—which can be quickly invoked without learning.

[0126] (2) Acquired learning methods

[0127] As another implementation method, the quantization value of the main control signal C can also be established autonomously through the online trial and feedback mechanism of the robot system in actual operation, that is, by adopting a post-learning method.

[0128] When the system faces a motion requirement that has never been recorded in historical experience, it first selects an existing value from the existing set of quantized values ​​of the master control signal C, or randomly generates a new quantized value within a preset safety boundary and includes it in C; then, based on the target motion unit (terminal T) and the required motion mode (M), a new CMT ternary combination is established in the W1 layer by adding associated neurons.

[0129] Following the basic method described above, each manipulation of any motion unit by the robot will form a ternary combination (C, M, T). These combinations are stored in the system in the form of the motion experience of the motion unit, constituting the robot's "control experience base".

[0130] However, the system records more than just isolated CMT combinations. During each operation, the system simultaneously collects execution result information through one or more of the following sensor feedback channels:

[0131] (a) Internal sensing information, including but not limited to joint torque feedback, tension changes, kinematic parameters such as position / velocity / acceleration;

[0132] (b) External perception information, including but not limited to changes in end pose obtained through visual sensors, and environmental feedback such as collisions or abnormal sounds obtained through auditory sensors.

[0133] The system correlates and compares the above feedback information with the current movement intention, ultimately forming a complete memory tuple of "Movement Combination (CMT) + Movement Effect (Feedback Perception)". This memory tuple not only records "what was done" but also "how well it was done", providing a basis for decision-making and optimization for subsequent similar movement needs.

[0134] When the system encounters a motion request with an existing motion record again, it makes a comprehensive judgment based on the current motion scenario, motion target, and past experiences stored in the historical memory: If the motion target is consistent with a past experience in the historical record, and that experience is marked as successful, the system prioritizes calling the CMT combination corresponding to that experience and its associated parameters and executes it directly; If the motion target is consistent with a failed experience in the historical record (e.g., voltage too high or too low), the system uses that failed experience as an error correction benchmark, adjusts the relevant parameters in a targeted manner (e.g., lowering or raising the voltage level, switching the working mode, etc.), generates a corrected new combination, and tries again. This correction process and its result are stored in the memory bank as a new motion experience.

[0135] If there are multiple identical or similar experiences in the past, the system will select the best or weighted merge based on the feedback effect of each experience in order to continuously optimize the accuracy and stability of subsequent control.

[0136] In this way, every operation of the system—whether successful or unsuccessful—is not isolated, but continuously accumulated in the control experience base as a memory tuple of "motion combination + motion effect". Successful examples are directly reused, while unsuccessful examples serve as a basis for error correction and guidance. Together, they constitute a closed loop of continuous learning and self-optimization for robot motion control. As the operating time increases and experience becomes richer, the system's response to known scenarios becomes increasingly accurate, and its exploration of unknown scenarios has clear historical references, gradually approaching the progressive evolutionary characteristics of "trial and error—correction—solidification—generalization" in the human brain's motor learning process.

[0137] It is particularly important to emphasize that in the aforementioned acquired learning method, all C values ​​generated through actual trials—regardless of whether the corresponding trials were successful or failed—are not pre-ordered in the W0 layer's storage space according to any numerical order or ascending / descending rules. In other words, the system records C values ​​following the single-element principle of "if it exists, call it directly; if it doesn't exist, store it in order": when a C value already exists in the storage space, the system directly calls that value; when a C value has not yet been recorded, the system stores it in the storage queue according to the chronological order of the actual actions, that is, the first one to occur is placed at the front of the queue, and the later ones are appended in sequence.

[0138] Throughout the process, the system does not sort or rearrange the stored C values ​​based on the magnitude of the absolute voltage value, the level of voltage, or any preset numerical indexing rules. This "store as it occurs, time-series priority" recording method not only reduces the computational overhead of storage management but also completely preserves the original time trajectory of motion exploration, providing natural data support for higher-level time-series analysis and behavior backtracking.

[0139] Furthermore, each C value in the storage queue corresponds to a unique and fixed storage location and a unique label. Once assigned, this location and label cannot be changed or moved during the system's lifetime.

[0140] The fundamental reason is that every basic element in layer W0 (including C values) is a globally shared resource within the entire control framework. Any C value may be referenced by one or more associated neurons (L neurons) in layer W1, forming cross-combination association mappings. If the storage location or label of a C value changes, all mapping relationships recorded by the associated neurons referencing that C value will become invalid due to invalid pointers, resulting in the complete destruction of the system's existing knowledge system.

[0141] Therefore, this invention adopts the element management principle of "once stored, fixed location, unique label, and non-transferable". New C values ​​are only allowed to be appended to the end of the queue, and existing C values ​​cannot be deleted, have their positions modified, or be renumbered. This mechanism ensures that all related neurons in layer W1 have permanent validity in referencing elements in layer W0, guaranteeing the stability and traceability of knowledge accumulation.

[0142] 2. Motion unit working mode group M

[0143] This embodiment configures four independent working modes to address the four core operating conditions of bionic joint motion: active drive, passive follow-up, posture maintenance, and precise positioning, covering all conventional working states of the robot joint:

[0144] U4: Passive compliant following mode, corresponding to the working condition where the main control drive signal is zero; in this mode, the motor has no active drive output, and only outputs a slight adaptive tension according to the external traction torque to achieve smooth following without jamming or slack.

[0145] U5: Speed-priority drive mode, suitable for scenarios involving rapid and continuous joint movements;

[0146] U6: Precise positioning mode, suitable for scenarios involving joint posture locking and point holding;

[0147] U7: Torque-adaptive drive mode, suitable for adaptive motion scenarios with load changes and flexible contact.

[0148] The four working modes are independent of each other and have fixed elements, which can be freely combined with any main control voltage level to achieve differentiated motion effects under the same voltage.

[0149] 3. Motion unit terminal group T (execution port element)

[0150] Terminal group T serves as the signal output interface for the upper-level neural logic to connect downwards to each independent motion unit. Each terminal in group T corresponds to an independent robot joint motion unit and is the unique basic element node of that unit in the control architecture.

[0151] The specific number of T groups is the same as the total number of independent motion units contained in the motion system. In the same motion system, each motion unit has a unique identifier in a T group, thereby ensuring that when the system activates a certain end number in a T group, the control signal can be accurately transmitted to the motion unit that uniquely corresponds to that end number.

[0152] In this embodiment, for ease of explanation, six independent output terminal elements are symbolically provided, which are respectively denoted as U8, U9, U10, U11, U12, and U13.

[0153] (II) Construction of the W1-associated neuron layer

[0154] The W0 layer only stores scattered and independent basic control elements, and cannot directly form effective action commands. This invention creatively sets up a W1 associated neuron layer, which is the core of action encapsulation and logic execution for a single motor unit. Each associated neuron is an independent neural excitation unit with a unique combination binding rule.

[0155] A single W1 layer associated neuron = a unique master control voltage element + a unique working mode element + a unique target motor unit terminal element. That is, each L-numbered neuron completely encapsulates a unique and unambiguous joint driving scheme. Once the neuron is excited and triggered, it will synchronously call its bound voltage parameters and working mode, and output control signals to the designated motor unit.

[0156] like Figure 4 As shown, this embodiment configures 7 independently associated neurons L10-L16 to form standardized action encapsulation logic, which can fully embody the core combinatorial innovation rule of this invention:

[0157] (1) Same voltage, different modes, driving different units (L10, L11)

[0158] L10: Binds the main control voltage U0 + U5 (speed mode) + U8 terminal; enables the U8 motion unit to work at a given speed under the U0 voltage;

[0159] L11: Binds the main control voltage U0 + U6 (position mode) + U9 terminal; enables the U9 motion unit to work at a given position under the U0 voltage;

[0160] Although both have identical main control voltage parameters, they exhibit completely different motion control effects due to their different operating modes and output terminals. This demonstrates that the present invention can achieve differentiated motion control under the same driving parameters through neuron combinations.

[0161] (2) Differentiating multi-unit reuse by independent neurons with the same parameters and patterns (L14, L16)

[0162] L14 and L16: Both are bound to U3 and U7 (torque adaptation mode) and corresponding terminal outputs;

[0163] The two sets of neurons have completely identical parameter configurations, but are independent neural nodes that can be triggered separately, individually, and multiple times to drive different motor units or multiple repetitions of the same unit under the same working conditions.

[0164] This design addresses the shortcomings of traditional control logic in achieving "batch reuse and independent triggering of standardized movements with the same parameters," providing core support for later homogenized multi-joint movements and repetitive bionic movements.

[0165] The remaining associated neurons L12, L13, and L15 respectively complete the cross-combination coverage of different voltage levels, positioning modes, and torque modes, achieving full coverage matching of basic parameters.

[0166] (III) Core control logic of this embodiment: fixed mapping mechanism of associated neurons

[0167] Figure 4 The core control logic of the architecture shown can be summarized as: an associated neuron, which represents a fixed driving rule.

[0168] Specifically, once each associated neuron (L10-L16) in layer W1 is established, a specific master control signal C, a specific working mode M, and a specific execution terminal T in layer W0 are permanently bound together in the form of a triplet (C, M, T) to form an immutable basic driving instruction.

[0169] Taking L14 as an example: This neuron locks U3 (a certain voltage level), U7 (a certain operating mode), and U12 (a certain motor unit terminal) together. Once L14 is activated, the system unconditionally drives the motor unit corresponding to U12 according to the U3 voltage and U7 mode—the output parameters do not change due to external interference, and they do not drift automatically over time. Once this rule is solidified, it is bound to the neuron itself, and its mapping relationship remains unchanged unless the system actively deletes or modifies the neuron.

[0170] In subsequent operation, this neuron can be activated by the excitation of any associated neuron with which it has a synaptic connection, thereby triggering the movement of the motor unit. The associated neuron can be a higher-level neuron (such as W2, W3) representing more complex movements, or it can be an input neuron representing sensory signals (such as vision, hearing, touch, torque feedback, etc.). Regardless of the activation pathway, once the neuron is activated, its output driving rules are completely consistent—that is, it always drives the same motor unit with the same voltage and the same pattern. This is precisely the core significance of this invention in "elementalizing" and "neuralizing" the control logic.

[0171] It should be noted that this embodiment ( Figure 4 The diagram only shows the fixed connections between the three basic elements (C, M, T) and the neurons associated with layer W1 in an open-loop manner, i.e., the direct driving pathways established through pre-setting or learning. As for the connections between the various sensory signals, environmental perception, and higher-level scheduling signals that indirectly activate L14 through the excitation of other neurons and layer W1, in... Figure 4 The diagrams are not drawn one by one—the reason being that showing all potential sensory input pathways, cross-layer connections, and feedback loops in the same figure would make the diagrams overly complex and difficult to clearly demonstrate the main logic of the basic control architecture. The aforementioned sensory signal access, cross-layer coupling, and closed-loop feedback mechanisms will be described in detail in subsequent embodiments, taking into account specific application scenarios.

[0172] (iv) Implementation of physical connections between neurons in computers

[0173] Figure 4 The control architecture shown is built based on the working model of biological neurons. The synaptic connections between neurons are dynamic and real-time during learning and memory processes—connection weights can be adjusted according to usage frequency, new associated neurons can be created, and invalid or redundant connections can be suppressed or marked. These dynamic characteristics dictate that this architecture cannot be directly implemented using fixed-wiring printed circuit boards; instead, it must rely on indexing mechanisms in computer storage systems to represent and store the mapping relationships between neurons.

[0174] Specifically, Figure 4 Each layer in the architecture maps to a list or storage region in a computer: layers W0, W1, and higher (W2, W3, etc.) each correspond to an independent storage structure. Each neuron in a layer uniquely corresponds to a storage unit, and the address of this storage unit serves as the unique identifier of that neuron within the entire control architecture. Each layer can be an independent physical memory, a contiguous physical address space within an independent physical memory, or a contiguous logical address space constructed from multiple physical memories—this invention does not limit this, and those skilled in the art can flexibly choose according to the actual hardware configuration.

[0175] In terms of hierarchical relationships, for ease of description and implementation, layer W0 is defined as the lowest level (basic element layer), and layers W1 and above are sequentially higher levels (association layers and orchestration layers). Within layer W0, each storage unit stores entities with independent physical or logical meaning—i.e., specific quantified values, working mode identifiers, terminal numbers, and other basic elements. Storage units in layers W1 and above do not store entities themselves, but rather the relationships between entities; their functional meaning is entirely defined and expressed by the connected W0 layer entities and the interconnected topology. In other words, neurons in layer W1 do not contain any independent semantic information; all their semantics derive from the combination of the W0 layer elements they are connected to.

[0176] According to the above provisions, in this invention, the stored content of each storage unit at all levels should at least express the following basic information:

[0177] (1) The entity or association represented by the storage unit: If the unit is located in the W0 layer, it stores the specific basic element entity, such as the quantized voltage value of the main control signal C, the type code of the working mode M, the port number of the motion unit terminal T, etc. in this embodiment; If the unit is located in the W1 layer or above, it does not store the entity itself, but records the labels and connection attributes of all other storage units directly connected to the unit. Its functional meaning is fully reflected in the connection relationship it establishes, and does not independently carry any value or meaning.

[0178] (2) The connection between the storage unit and the upper and lower storage units: that is, the upper storage unit number referenced by the storage unit and the record referenced by the lower storage unit form a complete bidirectional index chain to ensure that the cross-layer data path is traceable and verifiable.

[0179] (3) Specific relationships between all neurons connected to the storage unit: including but not limited to time relationships (such as sequential, synchronous or asynchronous), spatial relationships (such as physical adjacency on the same joint group or the same motion chain), and logical relationships (such as conditional triggering, enable mutual exclusion or priority constraints), providing a structured description basis for higher-level scheduling and orchestration.

[0180] (4) Inspired by the synaptic connection efficiency of biological neurons, a weight coefficient is assigned to each connection in this system. This weight coefficient is stored in a storage unit and is used to characterize the activation intensity, historical usage frequency, and current reliability of the corresponding connection path. The weight can dynamically increase with successful activation and failed attempts during system operation. As an optional implementation, the weight can also be decayed for connections that have not been stimulated for a long time, thereby achieving an adaptive adjustment capability similar to the "use it or lose it" principle of biological nervous systems.

[0181] It should be noted that the above four items are the basic requirements for the storage unit index information of this invention, but it does not mean that every storage unit in the system must simultaneously and completely contain all four items. Which fields are included and which are not included in the specific index structure can be flexibly selected based on the functional role of the storage unit in the control architecture—for example, neurons that do not participate in time-series scheduling do not need to include a time relationship field. Those skilled in the art, after reading this embodiment, can adaptively adjust the index fields according to specific application requirements, all of which fall within the scope of the technical teachings disclosed in this invention.

[0182] exist Figure 4 In the W0 layer, each storage unit not only stores a basic element (such as U0, U1, etc.), but also establishes a connection with higher-level neurons. For ease of description, the connection between a storage unit and a higher-level neuron is called the "output," and the index that specifies this connection is called the "output index." Correspondingly, the connection index with the lower-level unit is called the "input index."

[0183] In this embodiment, the basic form of the output index is (memory number, cell address, connection efficiency), and the input index is (memory number, cell address). Since there is only one synaptic connection efficiency value between two interconnected neurons, this connection efficiency can be expressed only in one index or stored simultaneously in both the input and output indices. The specific implementation method can be selected according to the actual engineering needs. If the system has a large number of backtracking operations on the motion loop, the connection efficiency can also be stored simultaneously in both the input and output indices to improve computational efficiency. In this embodiment and subsequent embodiments, for the sake of brevity, the connection efficiency is only set in the output index.

[0184] (v) Example of W0 layer index

[0185] The following is an example of the content of a cell in layer W0:

[0186] U0 = [[U0], [(W1, L10, R01), (W1, L11, R02)]]

[0187] U1 = [[U1], [(W1, L12, R11)]]

[0188] U3 = [[U3], [(W1, L14, R31), (W1, L15, R32), (W1, L16, R33)]]

[0189] Where [U0], [U1], and [U3] represent the specific content of the entities stored in the unit (i.e., the basic elements themselves); the list in the outer square brackets is the output index list of the unit, which records the connection relationship between the unit and each associated neuron in layer W1; (W1, L10, R01) indicates that there is a connection between the unit and the L10 neuron in layer W1, and R01 is the synaptic efficiency of the connection.

[0190] Taking U0 as an example: the entity content stored in this unit is U0 (a certain quantization level of the main control signal C), and its output index points to two related neurons L10 and L11 in layer W1, corresponding to two different control paths—L10 combines U0 with U4 and U8 into one driving rule, and L11 combines U0 with U5 and U9 into another driving rule. This shows that the same basic element can be shared by multiple high-level neurons, thus reflecting the element-sharing architecture feature of this invention.

[0191] The above example also applies to other basic element units in the W0 layer: the index format of group M (U4~U7) is exactly the same, the only difference being that the entity content stored in square brackets is the working mode type code; the index format of group T (U8~U13) is also exactly the same, the difference being that the entity content stored in square brackets is the terminal port number. All units in the entire W0 layer uniformly adopt the format of "entity content + output index list", and the structure between units is consistent, which facilitates programmatic processing.

[0192] (vi) Example of W1 level index

[0193] The index of each associated neuron in layer W1 adopts a three-level nested structure, designed to group the connected elements in layer W0 according to their functional roles. For example, consider layers L10, L11, and L12:

[0194] L10 = [[[(W0, U0), (W0, U5)], [(W0, U8)]], []]

[0195] L11 = [[[(W0, U0), (W0, U6)], [(W0, U9)]], []]

[0196] L12 = [[[(W0, U1), (W0, U5)], [(W0, U10)]], []]

[0197] The structure of the above index is explained as follows:

[0198] The first layer (outermost layer): The index is divided into two main parts: "input index" and "higher-level output reservation". The first square bracket is the input index, which records all input connections of this neuron from layer W0; the second empty square bracket is the higher-level output reservation, which is reserved for output connections of layer W2 and above. In this embodiment, it is temporarily empty.

[0199] The second layer (within the input index): The input index is further divided into two groups. The first group is the "input control group," which records the control elements that determine "how to drive" (including the quantization level of the main control signal C and the type code of the operating mode M); the second group is the "execution output group," which records the execution elements that determine "which" to drive (i.e., the port number of the motion unit terminal T). These two groups have fundamentally different functional roles: the first group constitutes a complete description of the drive command, while the second group indicates which specific motion unit the command should be sent to.

[0200] The third layer (within each group): consists of specific index entries. Each index points to a specific W0 layer basic element, including the element's level (W0) and cell number.

[0201] Taking L10 as an example: its input control group points to U0 and U5 in layer W0, indicating that the neuron receives U0 (a certain voltage quantization level) and U5 (a certain working mode) as control inputs; the execution output group points to U8 in layer W0, indicating that after the neuron is activated, it drives the motor unit terminal corresponding to U8. The three together constitute a complete driving rule - "drive the motor corresponding to terminal U8 with voltage U0 and mode U5".

[0202] L11 and L12 have identical structures; the only difference is the specific W0 unit label that the index entries in each segment point to, which corresponds to... Figure 4 The different connection relationships are shown in the diagram. L13 to L16 also follow the exact same three-part nested format, which will not be listed one by one.

[0203] (vii) Explanation of connection efficiency R value

[0204] In the index examples of W0 and W1 layers above, each index entry contains a connection efficiency parameter R (such as R01, R02, etc.). The R value represents the synaptic connection efficiency of a connection path. The label itself only serves as an identifier for that connection path and does not directly express the magnitude of the efficiency. Instead, its quantitative meaning is reflected through the specific numerical value corresponding to the label.

[0205] In actual operation, the R value is used to characterize the frequency or effectiveness of the connection path. It can be calculated simply by accumulating the number of times the path is activated, or other nonlinear algorithms can be used as needed. This invention only clarifies its meaning and does not specifically limit the specific calculation method.

[0206] (viii) Explanation of the flexibility of the index format

[0207] It should be noted that the above index formats—including the "output index list" of layer W0 and the "three-segment nested structure" of layer W1—are merely one of the exemplary implementations of this embodiment, used to illustrate how the four basic pieces of information of the aforementioned storage unit are implemented in a specific data structure. After reading this embodiment, those skilled in the art can use other equivalent data structures (such as hash tables, pointer arrays, linked lists, database tables, etc.) to express the same basic information according to specific application requirements, hardware environment, or programming language characteristics. All of these fall within the scope of the technical teachings disclosed in this invention and do not depart from the protection scope of this invention.

[0208] II. Example 2: Passive Following Control Mode Based on Torque Feedback

[0209] Human joint movement control relies on the coordinated work of antagonistic muscle groups. Taking the elbow joint as an example, the biceps brachii and triceps brachii form an antagonistic muscle pair: when the biceps brachii contracts to generate an active torque that drives joint flexion, the triceps brachii is not completely relaxed, but maintains moderate passive tension, providing physical damping for joint movement and preventing excessive swaying due to inertia; at the same time, this co-contraction mechanism ensures that the joint is in a controllable state at all times. When a rapid switch from flexion to extension is needed, the antagonistic muscle side is already in a pre-tensioned state, eliminating mechanical transmission gaps and making steering response almost instantaneous. This coordinated mode of "active side exerting force, antagonistic side pre-tensioning" is the fundamental reason why human movement combines flexibility and stability.

[0210] If the antagonist muscle is completely relaxed, the joint will be in an undamped, free-swinging state—the movement speed under external force is uncontrollable, and inertial swaying occurs after reaching the target position. Furthermore, when the antagonist muscle needs to switch to active contraction, it must first overcome mechanical backlash, causing a response delay. This is precisely the problem faced by traditional robot joints in power-off or zero-torque modes: when the motor rotor is in a free state, the joint cannot stably follow external forces, and re-establishing electromagnetic drive requires additional time, resulting in limited dynamic response.

[0211] This invention is in Figure 4Based on the active drive architecture shown, a passive following control mode that simulates the co-contraction mechanism of antagonistic muscles is further proposed. The core idea of ​​this mode is that when the master control signal C is zero (i.e., the system has no active drive command), the system automatically switches the control input from the master control signal C to the torque feedback signal FT, and calls the working mode U4 (passive compliant following mode), forming a mapping rule of associated neurons with the combination of "FT + U4 + T", which drives the motor to output a weak compensating torque that is in the same direction as the external traction torque but has a limited amplitude.

[0212] The function of this compensating torque is equivalent to the resting tension of biological muscles—it is insufficient to actively drive joint movement, but it is sufficient to eliminate the mechanical backlash and inertial slack of the motor rotor, so that the joint exhibits a smooth but not loose motion characteristic under external force drag, while keeping the motor in an electromagnetic hot standby state at all times, ensuring that it can intervene instantly and switch without delay when the active drive command is restored.

[0213] Figure 5 The diagram shown is a schematic representation of the control architecture of this embodiment. Figure 4 In comparison, its overall hierarchical structure (W0 basic element layer and W1 association layer) and index format remain unchanged. The only difference is that in the W0 layer, the torque feedback group FT replaces the main control signal group C. The FT group contains four quantization levels, U14 to U17, corresponding to different amplitude ranges of the torque sensor feedback signal. Its quantization method is the same as... Figure 4 Similar to Group C, it can be divided into uniform or non-linear partitions and supports both preset and online learning generation methods. U4 is the passive compliant following mode, which is the only enabled working mode in this embodiment. It is combined with each quantization level of the FT group and mapped to the corresponding motor unit terminals through the associated neurons of the W1 layer to form a complete passive following driving pathway.

[0214] like Figure 5 As shown, layer W1 is configured with seven interconnected neurons, L17 to L23. Each neuron binds a torque feedback quantization level, a U4 mode, and a motion unit terminal in the FT group, forming a complete passive following drive rule. Taking L17 as an example, this neuron locks U14 (a torque feedback value), U4 (passive compliant following mode), and U8 (a motion unit terminal) together: when the joint is subjected to external torque and the torque sensor feedback value reaches the quantization level corresponding to U14, L17 is activated. The system then outputs a weak drive current in U4 mode to the motor corresponding to U8, which is in the same direction as the external torque but with a limited amplitude. The amplitude of this current is limited to just enough to compensate for mechanical backlash and rotor inertia, without generating an active driving force exceeding the external torque. This ensures both the compliance of the following and the stability and response speed of the system.

[0215] because Figure 5 The architecture shown is Figure 4 Having the exact same hierarchical structure and index format, the index definitions of each storage unit in layers W0 and W1, the representation methods of input / output indexes, and the statistical methods of connection efficiency R values ​​are all completely consistent with those in Embodiment 1, and will not be repeated here. Those skilled in the art can determine without question the meaning of the index format description in Embodiment 1. Figure 5 The specific implementation of each neuron in the computer storage system.

[0216] It is worth noting that within the data architecture of this invention, each neuron (storage unit) in the basic element layer and each associated layer is unique. Specifically, in the above embodiments, this means that any newly added unit must be assigned an independent number, such as... Figure 5 The newly added U14, U15, U16, U17, and L17 to L23, etc.

[0217] The following explanation uses L17 as an example to illustrate the working characteristics in follow mode. The contents of cell L17 are as follows:

[0218] L17 = [[[(W0, U14), (W0, U4, R)], [(W0, U8)]], []]

[0219] When the torque feedback value is the same as the quantized value stored in the U14 unit, L17 is found through the output index of U14, and then the torque value is sent to the U8 terminal pointed to by its execution output group through the index of L17, completing the complete mapping from sensing to execution.

[0220] In actual operation, when the system reads a certain FT quantization value from the index and calculates the corresponding compensation torque, an adjustable gain coefficient α can be multiplied on the FT value at the program call level, by the upper-level decision logic, or the user control interface. This gain coefficient is not part of the index structure and is not permanently stored in the W0 or W1 layers. Instead, it serves as an instantaneous adjustment parameter at higher levels (W2 and above), dynamically applied to the read value when accessing the W1 layer index.

[0221] This design achieves a clear separation between the "lower-level data connection relationship" and the "higher-level control strategy," and has the following advantages:

[0222] (1) The index structure remains stable, and there is no need to add fields or change the format due to gain adjustment requirements. It is fully compatible with the storage architecture of Example 1.

[0223] (2) The following force can be flexibly adjusted according to different joints, working conditions, and scenarios. For example, in rehabilitation training scenarios, α can be lowered to make the joints have a greater damping feel to ensure safety; in precision assembly scenarios, α can be raised to achieve a more sensitive following response; in dual-arm collaborative handling scenarios, different α values ​​can be set for the active arm and the driven arm to optimize overall coordination. Each joint can be assigned an independent α value according to its role in the kinetic chain - for example, when the end effector needs to maintain a precise posture, the α value of the proximal joint can be appropriately reduced to reduce the transmission of external force disturbances to the end effector, while the α value of the distal joint is kept higher to ensure a sensitive response - none of the above adjustments require modification of the index content of the W0 or W1 layer.

[0224] (3) The gain coefficient is not part of the index structure and is not stored in the W0 or W1 layer. Instead, it is a dynamically configurable adjustment parameter in the higher layer (W2 layer and above). When accessing the W1 layer index, it is dynamically applied to the read value. It supports instantaneous adjustment and can also be continuously updated based on historical running effects.

[0225] From a biomimetic perspective, this mechanism corresponds to the gain modulation of the spinal reflex pathway by the cerebral cortex in biological motor control—the same feedback signal will be amplified or inhibited to different degrees by the central nervous system under different motor intentions or external environments to achieve differentiated motor responses. This invention achieves this biological feature in engineering by placing the gain coefficient outside the index structure and controlling it independently at a higher level.

[0226] Furthermore, although this embodiment uses torque feedback as the driving signal for passive following, those skilled in the art will understand that its core mechanism is also applicable to other feedback signals. For example, in a position deviation-driven following scenario, simply replacing FT with the position deviation signal ΔP while keeping the rest of the architecture unchanged will achieve the position following mode; in a speed deviation-driven following scenario, replacing FT with the speed deviation signal ΔV will achieve the speed following mode. The substitution of FT as the type of input signal does not change the core technical features such as the W0 / W1 hierarchical architecture, index format, and mapping mechanism, and all fall within the scope of the technical teachings disclosed in this invention.

[0227] III. Example 3: Formation of Continuous Actions

[0228] Any continuous movement of the human body can be divided into several sequentially connected independent time slices along the time axis. To simplify the description of motion logic and standardize the internal control logic of the system, this embodiment establishes a unified segmentation rule: within any single fixed time slice interval, the main control signals corresponding to all motion units participating in this movement are allowed to have equal or unequal values, but during the complete existence of the slice, the amplitude and intensity of each main control signal must remain constant; once any main control signal changes value, the moment of change is the end time of the current time slice, and also the start time of the next new time slice.

[0229] According to Example 1 (refer to) Figure 4 In layer W1, all associated neurons represent a specific movement pattern of a motor unit. Therefore, all motor units participating in active movement simultaneously within a time segment are first connected through the associated neurons in layer W2, forming a multi-joint simultaneous movement pattern across all time segments. Then, neurons in layer W3 are connected sequentially to represent movement modes in different time segments, ultimately piecing together a complete movement. A schematic diagram of the specific connection structure is shown below. Figure 6 As shown.

[0230] (a) Add time gap group S to layer W0

[0231] To enable flexible control of pause times between actions, this embodiment adds a set of time gap elements S (exemplarily denoted as U18, U19, U20, and U21) to the W0 layer to store pause time values ​​between adjacent time slices in a continuous action sequence. Each S element occupies an independent cell address and exists alongside existing elements such as C, M, T, and FT in the W0 layer.

[0232] The pause time values ​​stored in the S-element can be flexibly defined according to the actual application scenario—they can be counted by natural beats (such as "one beat" or "two beats"), by seconds (such as "0.5s" or "1.0s"), by minutes, or any other applicable time unit. This invention does not impose specific limitations on this. When no pause is needed between two adjacent time slices, the S-element with a value of zero (such as U18=0) or no S-element is referenced; when a pause of a specific duration is needed, the corresponding S-element is referenced (such as U19=0.5s, U20=1.0s). The specific values ​​of the S-element can be written or modified at any time in the W0 layer by the user or the system according to the actual working conditions, and its storage format is completely consistent with the C-group quantized voltage values.

[0233] At the biomimetic level, the S-element corresponds to the way organisms perceive pauses using heart rate or natural breathing rhythms as time references—organisms themselves do not possess precise millisecond-level timers; the determination of time intervals relies on endogenous rhythmic signals such as heart rate and respiratory rhythm, or on the synchronous input of external beat signals. This embodiment achieves the simulation of this biological characteristic in engineering by defining the S-element as a basic element that can be freely set by the user according to the beat or actual needs.

[0234] In order to make Figure 6 To maintain a clear structure and facilitate the demonstration of hierarchical connections in continuous actions, the elements of group S and the neurons in layer W3 are not shown as lines in the diagram. As the basic element group of layer W0, the reference relationships between group S and higher-level neurons will be specifically reflected in the index format of this embodiment.

[0235] (ii) Output index format of W1 layer neurons

[0236] Each associated neuron in layer W1 still uses the index structure defined in Example 1—the input index points to the basic elements (C, M, T, etc.) of layer W0, with the specific format shown in Examples 1 and 2; the output index points to the corresponding neuron in layer W2, with the format (layer identifier, unit number, connection efficiency R). Now, let's take... Figure 6 Taking the L10, L11, and L18 neurons as an example, the specific indexing format is explained below:

[0237] L10=[[],[(W2,L20,R100),(W2,L22,R101)]]

[0238] L11=[[],[(W2,L20,R110),(W2,L23,R111)]]

[0239] L18=[[],[(W2,L21,R180),(W2,L25,R181),(W2,L26,R182)]]

[0240] The index adopts a two-level nested list structure: the first part [] is the input index (in this example it is empty, indicating that this part has been defined in Examples 1 and 2 and will not be listed again here), and the second part is the output index list, which records all W2 layer neurons connected to the neuron and their corresponding synaptic connection efficiency R value.

[0241] As seen in the examples above, L10 simultaneously points to neurons L20 and L22 in layer W2, L11 simultaneously points to neurons L20 and L23, and L18 simultaneously points to neurons L21, L25, and L26. Each parenthesis (W2, Lxx, Rxxx) represents the target layer (W2), the target unit label (Lxx), and the synaptic connection efficiency (Rxxx) of that connection path, respectively. This index structure fully records which neurons in layer W2 can reuse a neuron in layer W1—the same W1 primitive can participate in the combination of multiple different time slices simultaneously. This is precisely the concrete manifestation of the "element sharing" architecture of this invention when expanding from layer W1 to layer W2.

[0242] (III) Indexing format of neurons in layer W2

[0243] Each neuron in layer W2 represents a complete time slice, and its input index records all concurrently acting neurons in layer W1 within that slice, along with their respective operating modes. Within the same time slice, all referenced neurons in layer W1 execute synchronously, each independently driving its corresponding motor unit according to a preset voltage and mode, without any sequential order or time difference. If only one motor unit participates in an action within a time slice, that neuron in layer W2 connects to only one corresponding neuron in layer W1.

[0244] The indexes of neurons in layer W2 use the same two-level nested list structure as layer W1: the first part is the input index, pointing to all neurons in layer W1 that make up this time slice; the second part is the output index, pointing to higher-level neurons in layer W3 that reference this time slice. Figure 6 Taking L20, L21, and L25 as examples:

[0245] L20=[[(W1,L10),(W1,L11)],[(W2,L30,R200),(W2,L32,R201)]]

[0246] L21=[[(W1,L12),(W1,L13),(W1,L18)],[(W2,L30,R210),(W2,L38,R211)]]

[0247] L25=[[(W1,L13),(W1,L18)],[(W2,L33,R250),(W2,L35,R251),(W2,L37,R252)]]

[0248] As can be seen from the example above: the first square bracket contains the input index, which only records the labels of the neurons in layer W1 that make up this time slice, in the format (W1, Lxx), and does not include the connection efficiency R value; the second square bracket contains the output index, which records the labels of the neurons in layer W3 that reference this time slice and their corresponding connection efficiency R value, in the format (W3, Lxx, Rxx).

[0249] The input index does not include R values ​​because the synaptic connection efficiency R value for the same connection path is the same physical quantity, uniformly recorded in the output index of the lower-level neurons. Specifically, in this embodiment: the connection efficiency between layers W1 and W2 is recorded in the output index of neurons in layer W1 (see the W1 layer output index format section); the connection efficiency between layers W2 and W3 is recorded in the output index of neurons in layer W2. When it is necessary to trace the R value of a connection path backward from a higher level, the lower-level neuron label recorded in the input index can be used to query the output index of that lower-level neuron. This design avoids redundant data storage and ensures the consistency of connection efficiency values.

[0250] If there are special requirements for reverse tracing speed in practical applications, an R-value field can be added to the input end as a redundant copy as needed. However, this is an optional feature at the engineering implementation level and does not change the core index architecture of this invention.

[0251] (iv) Indexing format of neurons in layer W3

[0252] The neurons in layer W3 connect the neurons in layer W2 in chronological order to form a complete action sequence. Therefore, each neuron in layer W3 needs to express at least the following three pieces of information: (1) which action segments (time slices of layer W2) make up the action sequence represented by this neuron; (2) the execution order of each action segment; and (3) the interval between action segments.

[0253] The index of neurons in layer W3 adopts the same two-level nested list structure as layers W1 and W2: the first part is the input index, which lists all neurons in layer W2 that constitute the action sequence in the execution order, and references the time interval element S of layer W0 at the position where a pause needs to be inserted (in this embodiment, it is exemplarily denoted as U18, U19, U20, and U21, and their specific stored values ​​correspond to different pause durations); the second part is the output index, which is reserved for higher levels (W4 and above), and is temporarily empty in this embodiment.

[0254] The specific indexing format is illustrated below using neurons L30, L31, L32, L36, and L37 in layer W3 as an example:

[0255] L30=[[(W2,L20),(W0,U18),(W2,L21)],[]]

[0256] L31=[[(W2,L22),(W0,U19),(W2,L23)],[]]

[0257] L32=[[(W2,L20),(W0,U20),(W2,L24)],[]]

[0258] L36=[[(W2,L27),(W0,U20),(W2,L28)],[]]

[0259] L37=[[(W2,L25),(W0,U18),(W2,L26),(W0,U21),(W2,L28)],[]]

[0260] As seen in the example above, the input index lists the elements sequentially according to the execution order, with the order indicating the execution sequence: elements appearing first are executed first, and those appearing later are executed later. Here, (W2, Lxx) represents the execution of a specific time slice; (W0, Uxx) indicates the insertion of a pause between adjacent time slices—Uxx points to the S groups of time gap elements stored in layer W0, the specific duration of which is determined by the value stored in that element. If Uxx=0, it means they are executed continuously without pause. The second set of square brackets [] is the output index, reserved for further arrangement at layers W4 and above (e.g., combining multiple complete action sequences into more complex compound actions, or repeating a set of actions, etc.). The principle remains the same, and this embodiment will not further expand upon it.

[0261] Taking L30 as an example: [(W2, L20), (W0, U18), (W2, L21)] means that the time slice represented by L20 is executed first, the duration represented by U18 is paused, and then the time slice represented by L21 is executed.

[0262] Taking L37 as an example: [(W2, L25), (W0, U18), (W2, L26), (W0, U21), (W2, L28)] means that L25 is executed first, U18 is paused, then L26 is executed, U21 is paused, and finally L28 is executed.

[0263] As can be seen from the examples above, the composition of the input index of the W3 layer neurons follows a unified rule: any number of W2 layer time slices and any number of S groups of time gap elements are alternately arranged in the actual execution order. The S elements are optional—they are inserted if a pause is needed, and if not, the pause time is set to 0 or not inserted; the position and duration of the pauses can be flexibly configured in the W0 layer according to actual needs. This serialized index format, which alternates between "action segments + time gaps," allows continuous actions of arbitrary complexity to be represented as a linear, time-ordered index sequence, completing the full expression of continuous action sequences within a unified data architecture.

[0264] It is worth noting that while adding time intervals to the neuron input indexes in layer W3, corresponding output indexes need to be added synchronously to the units such as U18 and U19 referenced in layer W0. The specific format is the same as in Examples 1 and 2, and will not be repeated here.

[0265] (v) Relationship between this embodiment and the foregoing embodiments

[0266] The neurons associated with layer W1 used in this embodiment are the action primitives L10-L16 defined in embodiment one. The newly added S group in layer W0 is listed alongside elements such as C, M, T, and FT, and is stored in layer W0 in the same format. All index formats from layer W1 to layer W3 are completely consistent with those in embodiments one and two, the only difference being that the reference target is expanded from layer W0 to layers W1 and W2. This structural continuity and consistency allows the system to simultaneously support single-joint active drive (embodiment one), passive following (embodiment two), and multi-joint continuous motion orchestration (this embodiment) under the same architecture. The three embodiments are compatible with each other and can work seamlessly together, forming the complete robot motion control system of this invention.

[0267] The hierarchical design, time slice division method, connection relationship and number of neurons in each layer adopted in this embodiment are all exemplary solutions set up to more clearly and intuitively describe the core idea of ​​the present invention, and are not the only limitation of the present invention.

[0268] In practical engineering applications, those skilled in the art can flexibly adjust the approach described in this embodiment based on specific working conditions, robot structure, motion accuracy requirements, response speed requirements, and other factors—including but not limited to: increasing or decreasing the number of neurons in each layer, adjusting the granularity of time slice division, increasing or decreasing the types of elements in layer W0, and expanding the number of layers W2 / W3 to accommodate more complex motion choreography needs. As long as the core architecture of "basic elements stored in layer W0, inter-layer relationships expressed by index, and high-level combinations forming complex actions" is followed, it falls within the scope of the technical teachings disclosed in this invention.

[0269] The flexibility of this invention is also reflected in the following: any newly added element type (such as control signals with different physical meanings, sensor feedback in different formats, pause intervals with different time units, etc.) can be stored as a new basic element in layer W0 and referenced by layer W1 and above in a unified format; any newly added action type or motion pattern can be incorporated into the existing hierarchical structure by adding associated neurons without affecting existing neurons and their connections. This "plug-and-play" extensibility makes this invention not only applicable to the robot motion control scenario described in this embodiment, but also generalizable to a wider range of control and orchestration tasks.

[0270] IV. Example 4: Multimodal Feedback Regulation at the Consciousness Level

[0271] In the human motor system, some movements are driven by reflex mechanisms that are innately fixed in neural circuits, such as the knee-jerk reflex and blinking, as well as follow-up movements driven by the active joints in the antagonistic muscle mechanism (corresponding to the passive following mode when the master control signal C=0 in Example 2). These types of movements do not require learning or conscious participation; they are automatically triggered by specific stimuli, have a fast response speed, high stability, and do not consume attentional resources.

[0272] However, the scenarios humans face are infinitely diverse, and the actions required are countless. Innate reflexes can only cover a very small number of specific situations—their main functions are self-protection, rapid obstacle avoidance, and maintaining the basic stability and smoothness of the motor system, far from sufficient to support the vast majority of purposeful motor behaviors in daily life.

[0273] What the human body inherits through genetics is a signal system and hardware architecture (skeleton, muscles, neural pathways) with the potential for movement, as well as the basic interface between actuators and control signals. The task of learning after birth is to build upon this foundation, first establishing the ability to flexibly manipulate each joint—enabling each joint to move independently as needed, and coordinating multiple joints to complete complex movements when required.

[0274] Since all movements are composed of joints combined in different ways, there are inevitably many commonalities between different movements: the same joint can serve a variety of different movement needs; the movement function of a joint is not designed for a specific movement, but rather is a fundamental resource serving the overall motor function. Therefore, the value of the experience in manipulating each joint accumulated during learning is not limited to the specific movement currently learned, but can be transferred to any other movement involving that joint. The commonalities between movements allow past learning outcomes to continue to influence future new movements—this is the fundamental reason why motor skills can be continuously accumulated and expanded.

[0275] The formation of motor skills requires a complete process of trial and error, correction, solidification, and recall. When learning a new movement, the system, guided by the conscious level, tries different joint combinations and control parameters, judging the effect of the movement based on the sensory signals (visual, tactile, proprioceptive, etc.) fed back during execution, recording successful paths and marking failed paths. Through repeated attempts, the most effective execution plan is gradually selected. However, trial and error alone is not enough to form a skill—each successful attempt must have its corresponding movement parameters and sensory feedback stored in the form of a "sensory-motor association," forming a recallable memory. Only after solidification can successful experiences be recalled at any time when needed, and the movement can transform from "conscious exploration" to "skillful automatic execution."

[0276] This embodiment, based on the "basic elements → action primitives → time slices → action sequences" framework established in the aforementioned embodiments one to three, focuses on demonstrating the inseparable and close connection between perceptual signals fed back through various senses and actions.

[0277] Figure 7 The diagram shown is a schematic structural diagram of this embodiment. This embodiment adopts a three-layer functional schematic architecture, denoted as WX, WY, and WZ. It should be noted that WX, WY, and WZ are only logical groups and are not actual physical layer labels such as W0, W1, and W2 in the data architecture of this invention. This layering is only used to characterize the relative positional relationship of signal input and output, facilitating logical organization and indexing: if a neuron in a WY layer receives input from WX and its output is connected to WZ, then this neuron is arranged between the corresponding unit in WX and the corresponding unit in WZ in the actual network; a neuron in a WZ layer receives input from WY, and its actual network layer position is higher than that of the neuron corresponding to WY. This relative positional relationship is a schematic expression and is not an absolute limitation.

[0278] The spatial arrangement of neurons does not determine their function. Similar to physical neurons in a living organism, the functional significance of a neuron is not determined by its spatial location, but entirely by the objects it connects to. However, within this element index framework, neurons need to be assigned a definite logical positional hierarchy; only by clarifying this logical position can the input and output indexes be defined, enabling the index to directionally point to the target unit.

[0279] The WX layer contains two regions, defined by dashed boxes and labeled A and B, respectively.

[0280] Part A is the sensory memory area. This area gathers feedback information from all sensory channels and the memories formed within the system. Based on the "element sharing" data architecture described above, layer W0 is the basic element layer, storing the smallest identifiable information units at the conscious level. For example, the raw visual signals received by the brain are a large number of pixels, but the objects of conscious recognition are not isolated pixels, but rather identifiable image primitives obtained by integrating pixels, such as primitives, outlines, or graphic regions with color and texture features; the basic elements corresponding to auditory signals are not raw sampling points, but recognizable sound segments with pitch, rhythm, and timbre features.

[0281] The granularity of basic elements is not fixed but determined relative to the application scenario and cognitive level. In language processing scenarios, Chinese characters can serve as basic elements; for more refined analysis, strokes and radicals can also be used as basic elements. In robot motion control scenarios, the quantization level of the main control signal, the encoding of the working mode type, the terminal number of the motion unit, the quantization level of torque feedback, and the perceptual units that can be identified in each modality of vision, hearing, touch, and proprioception all belong to the basic elements defined in this scenario. When implementing this invention, those skilled in the art can set the division method of basic elements according to the specific application scenario, and such transformations are all within the scope of the technical teachings disclosed in this invention.

[0282] In higher levels above W0 (W1, W2, W3, etc.), each neuron represents object or sensory information with different granularities, obtained by integrating basic elements step by step. Taking visual perception as an example, a neuron can represent a needle, a needle eye, an apple, a cup, or an object; the same type of sensory information is distributed at different levels according to its complexity.

[0283] This embodiment does not distinguish the specific distribution of sensory memories at each level. Instead, all neurons related to sensation are uniformly grouped into Part A, which serves as the system's sensory memory bank. Regardless of whether a sensory memory is actually located at W0, W1, W3, or other higher levels, as long as it is stored in the system as a sensory memory, it is presented in Part A as a unified sensory element for the WY layer to access and reference.

[0284] Part B is the motion execution area. This area gathers all the motion schemes that the system has learned and mastered. Combining the aforementioned embodiments one to three: layer W1 stores the basic motion primitives of a single joint; layer W2 stores the synchronous motion combinations of multiple joints under the same time slice; layer W3 stores the continuous motion sequence formed by chaining them together in chronological order; higher layers correspond to more complex composite motions.

[0285] This embodiment no longer subdivides the actual hierarchical level of each action scheme within Part B. Whether it's a basic single-joint action (W1 level), a multi-joint synchronous action combination (W2 level), or a complete continuous action sequence (W3 level and above), all are uniformly categorized into Part B as independent entries within the action library that can be invoked at the consciousness level. Each action entry exists in Part B in the form of complete functional semantics, available for reference by the WY level as needed.

[0286] WY and WZ are two-level association layers. The neurons in these layers do not have inherent attribute differences. Their function is to associate and bind various motion-related elements at the lower levels according to predetermined rules, forming different levels of combination relationships, based on system requirements. The motion combinations represented by the WZ layer are composed of the associated elements of the WY layer; and the associated elements of the WY layer are constructed from the basic elements of the WX layer. The WY layer is exemplarily configured with twelve neurons, LY0 to LY11; the WZ layer is exemplarily configured with seven neurons, LZ0 to LZ6.

[0287] When faced with a completely new movement, the system has a general understanding of the joints to be controlled at the conscious level, and the range of joints involved in the movement can be roughly determined, but a precise control mapping relationship has not yet been established. Constrained by the new work scenario, joints that should participate in the movement may not be activated, or the speed, rotation angle, and termination position of the joint output may not meet expectations. Therefore, movement learning requires multiple trial-and-error iterations to gradually converge and form a proficient movement pattern; furthermore, to achieve the same behavioral goal, there is no single standard movement, and the system can have multiple alternative execution schemes.

[0288] Figure 7 In the middle section, A10 and A11 are conceptual markers at the goal-driven level. They can represent conceptual identifiers such as action names and sub-steps in compound actions, corresponding to action tasks that can be represented by language, such as "forehand shot," "raising a cup," and "layup." These conceptual markers serve as index entry points: when a specific action task needs to be performed, all historical memory pathways bound to it are activated through this conceptual marker.

[0289] Taking the initial execution of the movement target corresponding to A10 as an example: the system selects movement combination B0 through conscious decision-making by the brain. This mapping is achieved through the associated neuron LY5 in the WY layer, which is the associated neuron driving movement combination B0. During the entire execution of the movement, the system collects proprioceptive feedback, including proprioceptive signals such as muscle contraction, stretching, and tactile sensation on the body surface, exemplarily denoted as A0; at the same time, the external results of the movement execution receive feedback signals through the senses, denoted as A5, such as the visual feedback of the ball going out of bounds when hitting a tennis ball, and the auditory feedback generated by hitting the racket frame.

[0290] The quality of an action can be further evaluated by associating and binding corresponding signals through neurons. Based on this association, when the A0 prophetic signal reappears, the system can retrieve the historical execution results corresponding to that action.

[0291] At this point, sensory units A0 and A5 are activated simultaneously, becoming associated sensory units that correspond to motor combination B0. The rules for constructing associated memories in this architecture are as follows: memory associations can be established between synchronously activated element units; the establishment of these associations is regulated by the brain's conscious selection, and not all synchronously excited units will be indiscriminately solidified into long-term memories. In this example, the system establishes an association between A0 and A5 through neuron LY0 in layer WY; furthermore, it selects associated neuron LZ0 in layer WZ and establishes a higher-order association between LY0 and LY5, thereby forming a complete sensory-motor neural pathway for motor combination B0.

[0292] When the target action represented by A10 is performed again, since a solid memory association has not yet been formed from the previous action, the second attempt may either replicate the previous action or select a different action. According to this learning principle... Figure 7 A10 is simultaneously associated with LY5 to LY8, and can drive B0 to B3 respectively, exemplarily representing multiple sets of action execution results generated during multiple trial and error processes; these include actions with correct execution results, incorrect actions, and intermediate state actions with varying degrees of quality. Correspondingly, LY0 to LY3 represent the perceptual association units formed in each attempt, and further construct their respective perceptual-motor neural circuits through LZ0 to LZ3.

[0293] Driven by the A11 target, the system generates three motion paths, which drive B0, B2, and B4 respectively via LY9 to LY11, ultimately forming the perception-motion loops corresponding to LZ4 to LZ6. Compared to the trial-and-error process of A10, the actions performed by A11 reuse the motion combinations B0 and B2 already used by A10, meaning that B0 and B2 can be shared by motions under two different target tasks.

[0294] Assuming that the four action attempts corresponding to A10 occur before A11, the perceptual association neurons LY0 and LY2 corresponding to B0 and B2 already exist as internal elements of the system. When executing the action task of A11, the system does not need to rebuild entirely new perceptual association units; it can directly reuse and share the existing LY0 and LY2.

[0295] Therefore, this architecture achieves element sharing at both the physical structure and data architecture levels. Whether in a biological organism or a robotic system, physical hardware such as joints and motion units are naturally shared resources; however, traditional robotic solutions often encapsulate each set of movements as independent program modules. While the hardware is reused, true shared access and invocation at the data level are not possible. In this architecture, motion combinations and their associated sensory connections serve as element units that can be referenced by multiple tasks. Different target tasks can directly reuse existing elements without repeatedly generating redundant associated entities, achieving unified sharing of physical hardware and data architecture.

[0296] Once the system has accumulated a certain number of candidate motion loops, when executing the same target task again, the system can either explore entirely new motion schemes or call upon existing historical candidate paths. Taking target A10 as an example, the system already has four candidate paths, LY5, LY6, LY7, and LY8, driving B0, B1, B2, and B3 respectively. When the intention to execute task A10 is generated, LY5 through LY8 will be activated first.

[0297] After LY5 to LY8 are activated, they do not directly drive the corresponding motor combinations to output actions. Instead, the signals are transmitted upwards through the WY layer, where the brain's consciousness performs a comprehensive evaluation and final path selection. Taking LY5 as an example: after LY5 is excited, LZ0 is activated along the pathway; LZ0 further activates LY0 in reverse, and LY0 then awakens the corresponding sensory units A0 and A5. Here, A0 represents proprioceptive feedback, and A5 represents the external execution effect feedback corresponding to this action. By retrieving historically stored proprioceptive perception and execution effects, the system can learn the quality of past execution results for this path. Similar sensory retrospection processes occur simultaneously in the other candidate pathways, LY6, LY7, and LY8.

[0298] The system comprehensively compares the historical sensations and performance effects corresponding to multiple pathways and selects one pathway for execution. Assuming the pathway corresponding to LY5 is ultimately selected, after confirming that the historical effects represented by the feedback from A0 and A5 meet the task expectations, the brain consciously selects this pathway, amplifies LY5 to maintain its excitation, and then issues control commands to the motor combination B0, initiating the corresponding action output. The path selection for the target task A11 follows the exact same working mechanism.

[0299] When the system finally selects and executes the action plan corresponding to B0, B0 will feed back to the body's perception A0 during its operation, and at the same time generate a perception signal A5 representing the effect of the action execution. A0 and A5 are activated synchronously, further driving LY0 to generate excitement; the activation of LY0 then drives the activation of LZ0, and LZ0 further enhances the excitement level of LY5, thereby activating the entire complete loop of A0, A5→LY0→LZ0→LY5→B0.

[0300] Based on the synaptic plasticity mechanism of biological neural networks, the efficiency of synaptic transmission between neurons is enhanced when neurons within a circuit are repeatedly excited synchronously. Therefore, the efficiency of synaptic connections along the entire circuit gradually increases in the action path that is frequently selected, and the individual neurons in the pathway gain stronger competitive activation capabilities, making them more easily triggered by subsequent tasks. After multiple iterations, the probability of the B0 action path being selected by the system will continue to increase.

[0301] Through this screening mechanism, motion loops with historically good performance are more likely to be selected, while erroneous paths are avoided as much as possible. Once the LY5 path is selected for execution, the complete loop from sensing units A0 and A5, through LY0, LZ0, and LY5 all the way to motion combination B0 is activated as a whole; with each successful invocation, the synaptic transmission efficiency between units in this complete loop is synchronously enhanced.

[0302] Figure 7 The internal loops use lines of varying thickness to represent the differences in loop connection efficiency and memory strength; the entire motion loop A0-A5→LY0→LZ0→LY5→B0 is set to the thickest line, representing the highest memory strength of this loop. The motion loops corresponding to A11 also have differences in memory strength in the actual system, and the corresponding mechanism is also applicable, but they are not shown one by one in this diagram.

[0303] It should be noted that the improvement in circuit connection efficiency is not entirely determined by whether the action is executed correctly, but is also regulated by the degree of attention the brain pays to that circuit. Some incorrect actions, if they bring significant adverse consequences, can also form strong memories, thereby enabling the active avoidance of such actions.

[0304] More importantly, this invention is based on a data architecture that shares elements. Each motion unit serves as a basic motion element, and various complex movements are constructed by combining these basic motion elements to form higher-level elements of different complexity. These learned, mature motion elements are essentially the skillful combination and application of basic motion elements, namely the movement capabilities of each joint. A set of skilled movements is not an independent, closed-loop program module; the motion elements used in the movement are not exclusively occupied by a single movement, but can be openly reused for various movement scenarios.

[0305] This means that through long-term exercise learning, not only is the mastery of a specific set of movements achieved, but also the training of each movement element (joint) involved in that movement is completed. Therefore, as exercise learning experience continues to accumulate, the system can skillfully manage more and more movement elements, especially the control ability of dynamic balance posture is fully honed; relying on the element sharing mechanism, the trained movement elements can be directly transferred and reused, thereby continuously reducing the difficulty of learning new movements.

[0306] Therefore, the perceptual information generated during movement plays a crucial role in the effectiveness of movement execution. There is a one-to-many mapping relationship between the target concept and the specific movement execution; relying solely on the target concept itself is insufficient to output correct actions in different scenarios. However, there is a more precise correspondence between action and perceptual feedback. The system needs to rely on the perceptual information stored in memory to recall the matching movement loop. When athletes underperform in competition, they often describe it as "not feeling anything," which essentially means they cannot retrieve the matching perceptual memory in the corresponding scenario, thus making it difficult to output precise movements. The commonly held notion of "muscle memory" is a cognitive misconception; muscles, as the mechanism for executing movements, do not possess memory capabilities themselves; all memory is stored internally within the system.

[0307] It should be noted that the coordination of perception and movement does not rely on a single sensation to directly trigger a complete set of actions. During the execution of complex movements, the brain continuously captures various sensory signals and dynamically adjusts the movements. Since awareness can only focus on one type of sensory information at a time, when completing continuous complex movements, awareness constantly switches between different sensory channels, completing multiple segments of the perception-movement loop step by step.

[0308] Taking the process of a person grasping a cup as an example: Before the hand touches the cup, awareness primarily focuses on visual perception, judging the distance between the hand and the cup based on visual feedback, and driving the arm to extend towards the cup; when the hand approaches the target position, the system switches its focus, driving the fingers to make a grasping motion. If the hand misses, it is determined that the previous visual judgment of distance was inaccurate, which is a very common phenomenon in the early learning stage. Just like a toddler learning to grasp, or someone subjectively judging that they can hit a ball but actually missing, the initial perceptual judgment is not absolutely accurate and requires repeated attempts to accumulate experience.

[0309] Once the fingers touch the cup, the focus of awareness quickly shifts from visual to tactile perception. Based on tactile feedback, the grip strength is adjusted; then, combined with the perception of friction and whether relative slippage occurs, a secure grip is confirmed before lifting the cup. The entire grasping process involves a continuous switching of the type of perception focused on, creating a closed loop where perception triggers action, and action generates new perceptions, gradually completing the task.

[0310] The same applies to dynamic movements such as running. You need to drive your limbs to complete the stepping motion while maintaining your body balance at all times. Once you sense signs of imbalance during the stepping process, your awareness will switch your attention to the balance perception and drive compensatory movements such as arm swing and waist posture adjustment.

[0311] The familiar motor circuits formed through repeated reinforcement not only ensure the reliability of movement execution but also provide a foundation for routine movements to not rely on conscious control throughout. Especially for periodic repetitive movements like walking, conscious control does not need to intervene in every detail of each step, only playing a supervisory and control role; relying on the solidified sensory-motor circuits, proprioceptive feedback is sufficient to maintain continuous movement. When abnormal situations occur, conscious control then intervenes again to adjust the movement.

[0312] V. Summary and Conclusion

[0313] This invention proposes a biomimetic motion control architecture based on the interconnectedness and element sharing characteristics of neurons in the brain's neural network. Various motion-related information is stored in the form of basic elements, and the elements are bound through indexes to form reusable action primitives. Then, higher levels realize the temporal arrangement of primitives and sensory-motor association memory. New functions are implemented by incrementally supplementing basic elements and associated indexes, while the original storage and association relationships remain unchanged, ensuring that existing knowledge does not become invalid.

[0314] Unlike existing robots that treat complete actions as independent data packets, lack shareable primitives between actions, and require new actions to be developed from scratch, this invention allows complex actions to be broken down into combinations of existing primitives. Shared parts can be directly reused, while unique parts only need to be supplemented with a small number of primitives or indexes. Learning new actions is essentially a recombination of existing primitives, and the learned experience can be transferred to other actions containing the same primitives, thus building an experience network that can be accumulated and transferred.

[0315] The architecture's ability to imitate, learn, and generalize stems from the atomization of basic elements, global sharing of action primitives, diverse combinations and arrangements of primitives at higher levels, and a sensory-motor associative memory mechanism. The system's motion experience primarily comes from its own perceptual feedback and motion attempts. Through continuous trial and error, it iterative optimization, and gradual convergence, it obtains a motion scheme suitable for the scenario. This differs from the learning mode of existing robots that rely on externally imported datasets or pre-programmed complete motion sequences.

[0316] This architecture, at the functional level, mirrors the learning and regulatory characteristics of biological motor systems. The descriptions of consciousness and intention in the manual are merely biomimetic analogies; decision-making functions at the consciousness level are implemented through judgment logic and control processes within the program. The accompanying diagrams are schematic representations of logical principles and are not precise replicas of biological neurons. The digital system achieves functional equivalence through indexing.

[0317] The index format, unit numbering, synaptic connection efficiency R, and other parameters and arrangements are all schematic diagrams and not the only implementation scheme limited by this invention. In particular, Embodiment 4 only illustrates the association mechanism between perception and motion and does not limit the specific index format. Therefore, the standardized index structure shown in Embodiments 1 to 3 is only used to establish the core principle of "replacing biophysical synaptic connections with indexes": the index is used to accurately define the unit connection relationship between upper and lower levels, synaptic transmission efficiency, and the spatiotemporal association logic between various elements. Under the premise of following this core principle, those skilled in the art can independently adapt, optimize, and define specific index forms and parameter structures according to the specific engineering scenario requirements.

[0318] Based on this unified incremental expansion paradigm, the system can be extended to realize various functions such as lock-and-hold, force-position hybrid control, impedance adaptive adjustment, multimodal sensing fusion, and multi-machine collaboration. The number of neurons, quantization levels, and parameters listed in the embodiments are exemplary settings. Various modifications and substitutions made by those skilled in the art, while retaining the core mechanisms of element atomic storage, index-based association establishment, hierarchical incremental expansion, and shared and reusable primitives, all fall within the protection scope of this invention. The protection scope of this invention should be determined by the content of the claims.

Claims

1. A general biomimetic motion control model based on a data architecture for human brain neuron connections and element sharing, characterized in that, Includes servo actuators, sensing units, and controllers; The controller constructs a shared data architecture for elements and stores the underlying basic element set and related layer data under this architecture; The underlying basic element set is used to decompose motion control parameters, port information, and multimodal perception information into the smallest identifiable independent basic elements for independent storage. The basic elements have two sources of generation: the innate source is generated based on the hardware structure parameters during system initialization, and the acquired source is generated autonomously by the sensor unit through real-time perception feedback during system operation. Each basic element is assigned a unique and fixed storage label. Once the basic element is stored, it is not copied or migrated. New basic elements are only appended to the end of the storage queue. The association layer contains multiple levels of logical association neurons. Each level of logical association neuron generates a corresponding high-level element, and all elements at all levels support global sharing. The higher-level part of the association layer can serve as an action orchestration layer, used to combine multiple logically associated neurons according to temporal or logical relationships to construct action sequences; logically associated neurons bind two or more elements through index references to form action primitives; complex movements are constructed by combining elements at different levels through the interconnection of logically associated neurons at each level; The index includes at least a hierarchy identifier, a unit number, and a synaptic connection efficiency parameter, which is used to characterize the connection relationship between neurons. It can also record the temporal, spatial, or logical relationship between neurons. All neural connections, action combinations, and logical relationships are implemented by binding to the upper-level index. The controller is configured to: on the one hand, dynamically establish a perception-motor correlation loop between perceptual elements and motor elements under the guidance of conscious decision-making logic; it can invoke and execute actions through conceptual identifiers or stimulate corresponding action loops based on sensory memory; on the other hand, it calls action primitives to generate driving instructions and outputs them to the servo actuator, and receives feedback from the sensing unit, adjusting the synaptic connection efficiency parameter of the index according to the execution effect; the synaptic connection efficiency is used as the basis for conscious decision-making logic to filter motor loops, and the essence of things or actions is determined by the connection relationship between neurons.

2. The general-purpose bionic motion control model according to claim 1, characterized in that, The synaptic connection efficiency parameter of the index can be configured in the following locations: it can be set only in the output index; or, for reverse tracing requirements, the synaptic connection efficiency parameter can be configured at both the input and output indices.

3. The general-purpose bionic motion control model according to claim 1, characterized in that, The synaptic connection efficiency parameter can dynamically increase with the synchronous activation of the corresponding connection path, regardless of the correctness of the execution result of a single activation. As an optional implementation, synaptic connection efficiency decay can also be performed on connection paths that have not been stimulated for a long time, simulating the adaptive regulation characteristics of the biological nervous system of use and disuse. The higher the synaptic connection efficiency, the deeper the memory of the corresponding motor circuit, and the greater the probability of being competitively selected for execution.

4. The general-purpose bionic motion control model according to claim 1, characterized in that, The action unit includes an active driving unit and a passive compliant following unit; the passive compliant following unit is formed by binding torque feedback basic elements, a preset passive working mode, and execution port elements; when the main control driving signal is withdrawn, the controller triggers the passive compliant following output with limited amplitude compensation torque based on the torque feedback of the sensing unit.

5. The general-purpose bionic motion control model according to claim 1, characterized in that, The same element can be shared and referenced by multiple logically related neurons; the same logically related neuron can also be shared and referenced by multiple action sequences at higher levels of the related layer, relying on element sharing to realize the transfer and reuse of motion experience between different actions.

6. The general bionic motion control model according to claim 1, characterized in that, It also includes a feedback control layer; the feedback control layer sets up consciousness-associated neurons, which bind sensory elements in the sensory memory area with action plans in the action execution area through indexing, establishing a perception-motor association mapping; the same motion target can correspond to multiple perception-motor paths, and each path competes for the best execution based on synaptic connection efficiency parameters combined with execution result labels; as long as the perception-motor association path is synchronously activated, the synaptic connection efficiency of the corresponding path is enhanced regardless of whether the current action execution result is correct or incorrect; after the action is executed, a success label or a failure label is generated for the path based on the matching result of real-time sensory feedback and historical sensory memory; during the path competition and selection process, paths with failure labels, even if they have high synaptic connection efficiency, will be avoided, completing trial and error iteration.

7. The general-purpose bionic motion control model according to claim 1, characterized in that, When constructing continuous motion, the continuous motion is divided into several time slices along the time axis. The control signals corresponding to the motion units in a single time slice remain constant. Logically related neurons that work synchronously in the same time slice are combined, and then the time slices are connected in series according to the sequence of motion. Time gap elements from the underlying basic element set can be inserted between adjacent time slices to achieve pause control.

8. A control implementation method based on the general biomimetic motion control model according to any one of claims 1-7, characterized in that, Includes the following steps: (1) Decompose motion control parameters, port information, and multimodal sensing information into the smallest independent basic elements and store them in the underlying basic element set; the basic elements include the innate elements preset by the system hardware parameters, as well as the acquired elements obtained by the system operation through autonomous accumulation by sensing feedback; each basic element is assigned a unique fixed storage label, the basic element itself is not copied or migrated, and newly added basic elements are only appended to the end of the storage queue. (2) Construct multiple levels of logical connection neurons in the connection layer. Each level of logical connection neuron generates the corresponding high-level elements. All level elements support global sharing. Logical neurons bind two or more elements through index references to form action primitives; the higher-level part of the association layer combines multiple logically associated neurons according to temporal or logical relationships to construct action sequences; all neural connections and action combinations are realized by upper-level index binding; the index includes hierarchical identifiers, unit numbers, and synaptic connection efficiency parameters, recording the connections between neurons and temporal, spatial, and logical relationships; the essence of things or actions is determined by the connection relationships between neurons, and synaptic connection efficiency is used for loop filtering. (3) When a motion execution request is received, the controller can either call the action through conceptual identification or rely on sensory memory to stimulate the corresponding perception-motor association loop, and call the corresponding logical association neurons from the high level to the low level along the index to generate driving instructions to drive the servo actuator to complete the motion output; (4) Collect feedback information from the sensing unit, dynamically establish a perception-motor association loop under the guidance of the consciousness decision-making logic, and dynamically adjust the synaptic connection efficiency parameters of the index according to the motor execution effect to complete the solidification and transfer of motor experience.

9. The control implementation method according to claim 8, characterized in that, The steps for constructing continuous motion include: dividing the continuous motion into several time slices along the time axis, keeping the control signals of the motion units within a single time slice constant; combining logically related neurons that work synchronously within the same time slice; then connecting the various time slices in sequence according to the order of the motions, and inserting time gap elements between adjacent time slices as needed to achieve pause control.

Images