A training method and system for a quadruped robot walking controller

Abstract

This invention discloses a training method and system for a quadruped robot walking controller, belonging to the field of robot control technology. Specifically, it includes: collecting motion data from various quadruped robots in different walking scenarios, extracting common features related to walking control, and storing these common features in a common feature library; based on the common feature library, analyzing an existing adaptive walking controller trained for a specific quadruped robot, separating the control parameters and control strategies in the controller; considering the differences in the body characteristics of different quadruped robots, selecting key body characteristic parameters, and establishing a control parameter mapping model; based on the common features in the common feature library and the separated control strategies, constructing a control strategy transfer model; and integrating the initial control parameter values ​​of the new quadruped robot generated through the control parameter mapping model with the control strategy adapted through the control strategy transfer model to form a walking controller for the new quadruped robot.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, specifically to a training method and system for a quadruped robot walking controller. Background Technology

[0002] Existing training methods for quadruped robot walking controllers primarily rely on deep reinforcement learning (DRL) and adaptive control techniques. These methods simulate robot movement and optimize control strategies by constructing simulated environments. For example, a two-stage training framework integrates proprioception and contact force estimation to achieve robust walking in complex terrains; the Behavior Diversity Adjustment (MoB) method encodes a family of structured motion strategies, enabling a single strategy to adapt to different environmental requirements. Furthermore, bio-inspired DRL frameworks significantly improve robot adaptability in complex terrains through gait transition strategies and pseudo-gait procedural memory. These techniques have yielded significant results on specific robot bodies. For instance, the Unitree A1 robot achieved omnidirectional movement and traversal of small obstacles through training solely based on proprioception, while the Walk These Ways framework, trained on flat ground, can be directly applied to unseen environments such as stairs and slippery surfaces.

[0003] However, existing adaptive walking controllers trained on specific quadruped robot bodies (with specific mass distribution, size, joint characteristics, etc.) have very low or almost no adaptability to other quadruped robots of the same type with significantly different geometries or dynamic characteristics (such as leg length, weight variations, and motor torque differences). This means that for different individual robots or changes in body design, a time-consuming and resource-intensive training process (including remodeling, simulation, and learning strategies) often needs to be carried out from scratch. Existing adaptive methods lack effective mechanisms to mine, encapsulate, and transfer common walking control knowledge and structurally invariant characteristics among different individual robots, which greatly limits their universality and deployability. Summary of the Invention

[0004] The purpose of this invention is to provide a training method and system for a quadruped robot walking controller, solving the following technical problems:

[0005] Existing adaptive walking controllers based on training on specific quadruped robot bodies have very low or almost no adaptability to other quadruped robots of the same type with significantly different geometries or dynamic characteristics.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A method for training a quadruped robot walking controller includes:

[0008] S1: Collect motion data of various quadruped robots in various walking scenarios, extract common features related to walking control, including gait cycle features, joint motion trajectory features, and body posture balance features, and store the common features in a common feature library;

[0009] S2: Based on the common feature library, the adaptive walking controller trained for a specific quadruped robot is analyzed, and the control parameters and control strategies in the controller are separated.

[0010] S3: Based on the differences in the body characteristics of different quadruped robots, key body characteristic parameters are selected, including leg length, weight, and motor torque. By analyzing the correspondence between key body characteristic parameters and control parameters, a control parameter mapping model is established.

[0011] S4: Based on the common features in the common feature library and the separated control strategies, construct a control strategy transfer model, and train the control strategy transfer model using walking data from various quadruped robots.

[0012] S5: The initial values ​​of the control parameters of the new quadruped robot generated by the control parameter mapping model are integrated with the control strategy adapted by the control strategy transfer model to form a walking controller for the new quadruped robot.

[0013] As a further aspect of the present invention: in step S2, the control parameters are specific values ​​related to the robot's body characteristic parameters, and the control strategy is the logic and rules for walking control based on common features.

[0014] As a further aspect of the present invention: in step S2, the process of separating the control parameters and control strategy in the controller is as follows:

[0015] The existing adaptive walking controller is functionally decomposed into a gait planning module, a joint control module, and a balance adjustment module.

[0016] The inputs of each functional module are matched with common features in the common feature library to determine the types of common features that the module depends on for operation; among them, the gait planning module is associated with gait cycle features, the joint control module is associated with joint motion trajectory features, and the balance adjustment module is associated with body posture balance features.

[0017] The gait planning module's periodic generation logic is transformed into a conditional rule expression based on gait periodic characteristics; the balance adjustment module's posture correction logic is transformed into a functional relationship expression based on body posture balance characteristics; and the joint control module's trajectory tracking logic is transformed into a parameterized equation based on joint motion trajectory characteristics, forming a symbolic rule set.

[0018] Three sets of quadruped robot samples with different body characteristic parameters were selected, input into the controller, and the output changes of each module were recorded. Through comparative analysis, the numerical variables in the symbolic rule set that change with the body characteristic parameters were identified and marked as candidate control parameters.

[0019] The logical structure in the symbolic rule set that does not change with the robot's ontological characteristic parameters is defined as the core of the control strategy; the numerical variables in the candidate control parameters that have a monotonic mapping relationship with the robot's ontological characteristic parameters are defined as control parameters, thus completing the separation of control strategy and control parameters.

[0020] As a further aspect of the present invention: the process of establishing a control parameter mapping model by analyzing the correspondence between key ontological characteristic parameters and control parameters is as follows:

[0021] Quadruped robots with different leg lengths, weights, and motor torques were selected as samples to form a sample set of key body characteristic parameters;

[0022] For each quadruped robot in the sample set, all its control parameters under the standard walking scenario are obtained, and a control parameter dataset corresponding one-to-one with the key body characteristic parameter sample set is constructed.

[0023] Using key ontological characteristic parameters as row vectors and control parameters as column vectors, the Pearson correlation coefficient between any two parameters is calculated to form a parameter correlation matrix. The matrix element values ​​represent the degree of linear correlation between the corresponding two parameters.

[0024] Based on the parameter correlation matrix, a hierarchical clustering algorithm is used to group the control parameters. Control parameters with high correlation coefficients with the same key ontological characteristic parameter are grouped into a parameter influence cluster, and each cluster corresponds to the influence range of a key ontological characteristic parameter.

[0025] For each parameter-affected cluster, the corresponding key ontological characteristic parameters are used as inputs, and the control parameters within the cluster are used as outputs. The gradient boosting regression algorithm is used to train the parameter mapping sub-model.

[0026] All parameter mapping sub-models are chained together and integrated to construct a control parameter mapping model.

[0027] As a further aspect of the present invention: the method for constructing the control strategy transfer model is as follows:

[0028] A strategy feature encoding module is constructed to convert gait cycle features, joint motion trajectory features, and body posture balance features from the common feature library into a fixed-dimensional feature matrix; a strategy logic conversion layer is built, adopting a gated loop unit structure, with the input being a symbolic sequence of the separated control strategy logic expressions; a ontological characteristic modulation layer is set up to receive leg length, weight, and motor torque parameters and generate modulation coefficients; a strategy output decoding layer is constructed to convert the processed signals into an executable control strategy instruction set. The modules are connected sequentially according to the data flow to form a complete control strategy migration model architecture.

[0029] As a further aspect of the present invention: the process of training the control strategy transfer model using walking data from various quadruped robots is as follows:

[0030] When preparing the training dataset, walking data of various quadruped robots in flat ground, slope, and staircase scenarios were collected. Common feature matrices, ontological characteristic parameters, and the actual control strategy instruction set executed were extracted for each data point. A joint loss function was designed, constructing a strategy logic deviation loss term using edit distance and a common feature consistency loss term using cosine similarity, which were then added with 1:1 weights to form the joint loss function. Phased training was performed. First, the flat ground scenario data from the training dataset was used for training with gradient descent, with a preset base number of iterations. Then, slope and staircase scenario data were added for further training, with half the base number of iterations. Each iteration used the validation dataset to calculate the loss value and record the model parameters. Model adaptation fine-tuning was performed, selecting the model parameters with the smallest loss value from the validation dataset as initial parameters. Fine-tuning was performed individually for each robot sample, inputting its ontological characteristic parameters and common feature matrix, using the deviation between the output and the actual control strategy instruction set as the fine-tuning loss. The ontological characteristic modulation layer parameters were updated through a single iteration to complete model training.

[0031] As a further aspect of the present invention: the process of forming the walking controller for the new quadruped robot is as follows:

[0032] The control strategy after adaptation by the control strategy transfer model is analyzed, all logical nodes that require input control parameters are extracted, and the parameter types corresponding to each node are recorded; the initial values ​​of control parameters generated by the control parameter mapping model are classified by type, and a parameter strategy association index table is constructed.

[0033] Design an interface adaptation layer whose input is a set of initial control parameter values ​​and an adapted control strategy, and whose output is the strategy execution sequence after parameter embedding.

[0034] Based on the parameter strategy association index table, the initial values ​​of control parameters are injected one by one into the corresponding logical nodes of the control strategy through the interface adaptation layer. During the injection process, the logical structure of the control strategy remains unchanged, and only the parameter placeholders in the nodes are replaced to form a complete strategy expression with specific parameters.

[0035] Input the complete strategy expression into the simulation environment. The simulation environment has a built-in virtual model of a quadruped robot, whose body characteristic parameters are consistent with the new quadruped robot. Run the simulation environment and monitor whether there are any abnormalities such as parameter out-of-bounds or logical conflicts during the execution of the strategy expression. Record the location of the logical node where the abnormality occurs.

[0036] For any anomalies encountered during verification, the numerical range of the initial values ​​of the corresponding control parameters is adjusted through the interface adaptation layer, and the parameter injection operation is re-executed until no anomalies occur in the simulation environment. The complete policy expression that has passed verification is converted into code format that the controller can execute, forming a walking controller for the new quadruped robot.

[0037] A quadruped robot walking controller training system, comprising:

[0038] Common feature acquisition and database construction module: Collects motion data of various quadruped robots, extracts common features of gait cycle and joint motion trajectory, and stores them in the common feature database;

[0039] Controller parameter and strategy separation module: Based on the common feature library, the adaptive walking controller of a specific quadruped robot is analyzed and the control parameters and control strategy are separated.

[0040] Mapping model construction module: Based on the differences in the body characteristics of different quadruped robots, key parameters are selected, their correspondence with control parameters is analyzed, and a mapping model is established.

[0041] Control strategy transfer model construction module: Based on the common features in the common feature library and the separated control strategies, a control strategy transfer model is constructed and trained using walking data from various quadruped robots;

[0042] The walking controller integration module integrates the initial values ​​of the new quadruped robot control parameters generated by the control parameter mapping model with the adapted control strategy to form a walking controller;

[0043] Walking controller optimization and adaptation module: Optimize the initial controller using the actual walking data of the new quadruped robot, adjust the control parameters and fine-tune the strategy to obtain a suitable walking controller.

[0044] The beneficial effects of this invention are:

[0045] This invention significantly improves the controller's adaptability to robots with different body characteristics by systematically mining common knowledge in quadruped robot walking control and establishing a transfer mechanism, effectively solving the problem of traditional methods requiring retraining from scratch. On one hand, by extracting common features such as gait cycles, joint trajectories, and body balance and constructing a feature library, the control strategy and control parameters are separated, allowing the control strategy to be reused across different robots. Simultaneously, the control parameter mapping model, by learning the correlation between body characteristics and parameters, can quickly generate suitable parameters for new robots, eliminating the time-consuming process of remodeling and simulating for individual robots and significantly reducing resource investment.

[0046] On the other hand, this invention enhances the universality and deployability of the controller through a multi-scenario training and simulation verification mechanism for the control strategy transfer model. The strategy transfer model, based on common features and symbolic rules, combines an ontology-based modulation layer to adapt to different robots. Through phased training from flat ground to complex terrain, it ensures the stability of the strategy in diverse scenarios. Simulation environment verification and parameter fine-tuning further correct errors in the mapping model, allowing the initial controller to be directly applied to new robots without significant adjustments. This approach retains the common knowledge of walking control for different robots while adapting to individual differences through parameter mapping and fine-tuning, completely breaking the dependence of traditional controllers on specific ontologans and significantly improving deployment efficiency. Attached Figure Description

[0047] The invention will now be further described with reference to the accompanying drawings.

[0048] Figure 1 This is a flowchart illustrating a training method for a quadruped robot walking controller according to the present invention.

[0049] Figure 2 This is a schematic diagram of the structure of a quadruped robot walking controller training system according to the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Please see Figure 1 As shown, the present invention is a training method for a quadruped robot walking controller.

[0052] S1: Collect motion data of quadruped robots with different geometric structures and dynamic characteristics in various walking scenarios, extract common features related to walking control, including gait cycle features, joint motion trajectory features, and body posture balance features, and store these common features in a common feature library.

[0053] Specifically, the collected motion data from various quadruped robots underwent denoising, time synchronization, and standardization to remove errors caused by sensor noise and data acquisition delays, and to unify the data sampling frequency and timestamp format. Based on foot force sensor data, the ground contact and lift-off times of each leg were identified, and the duration of the support phase and swing phase of a single leg was calculated to determine the time span of the complete gait cycle. By comparing the gait cycle structure of different robots under the same walking mode, common parameters unrelated to the robot's inherent characteristics, such as the proportion of the support phase and the phase difference between legs, were extracted within the gait cycle. The time axis of the sequence of joint angle changes over time was normalized, mapping joint motion trajectories of different cycle lengths to the same time dimension. A dynamic time warping algorithm was used to align the motion trajectories of the same joints from different robots, calculating the curvature extrema and slope of the trajectory curves, and extracting the key inflection points and trajectory morphology features of joint motion as common features. The roll angle, pitch angle, and rate of change of the robot body are calculated using inertial measurement unit (IMU) data. A support polygon is constructed by combining this with foot position data. The relative offset between the projected coordinates of the body's center of gravity within the support polygon and the polygon's geometric center is calculated. The attitude angle stability threshold range and the temporal variation pattern of the center of gravity offset are extracted as common features. The extracted gait cycle features, joint motion trajectory features, and body posture balance features are dimensionally aligned. A contrastive learning network is used to learn the mapping relationship between different robot feature spaces. Feature parameters with a consistent distribution pattern across all robot samples are selected to form a set of common features related to walking control.

[0054] S2: Based on the common feature library, the existing adaptive walking controller trained for a specific quadruped robot is analyzed, and the control parameters and control strategies in the controller are separated. The control parameters are specific values ​​related to the characteristics of the robot body, and the control strategies are the logic and rules for walking control based on common features.

[0055] S3: To address the differences in the body characteristics of different quadruped robots, key body characteristic parameters are selected, including leg length, weight, and motor torque. By analyzing the correspondence between these parameters and control parameters, a control parameter mapping model is established. This model can generate corresponding initial values ​​of control parameters based on the input body characteristic parameters of the new quadruped robot.

[0056] S4: Based on the common features in the common feature library and the separated control strategies, a control strategy transfer model is constructed. The model is trained using walking data from various quadruped robots, enabling the model to learn the transfer rules of control strategies between different robots and ensuring that the control strategies can be adapted to different quadruped robots.

[0057] S5: The initial values ​​of the control parameters of the new quadruped robot generated by the control parameter mapping model are integrated with the control strategy adapted by the control strategy transfer model to form a walking controller for the new quadruped robot.

[0058] In step S2, the process of separating the control parameters and control strategy in the controller is as follows:

[0059] The existing adaptive walking controller is functionally decomposed into a gait planning module, a joint control module, and a balance adjustment module. By observing the complete workflow of the controller when implementing walking functions, we can identify the parts that perform different core tasks: For example, the part responsible for determining the robot's leg-stepping sequence, the time interval between each step, and the stride length is the gait planning module; the part responsible for controlling how each joint, such as the hip and knee joints, rotates (e.g., rotation angle and speed) to complete the stepping action is the joint control module; and the part responsible for adjusting the movement of each joint to maintain body stability when the robot tilts or encounters disturbances is the balance adjustment module. The principle behind this decomposition is that the walking control of a quadruped robot essentially consists of three continuous and independent tasks: "planning the walking rhythm," "executing joint movements," and "maintaining body balance." Each task corresponds to a functional module, facilitating the individual analysis of its internal logic.

[0060] The inputs of each functional module are matched with common features in the common feature library to determine the types of common features that the module depends on for operation. Specifically, the gait planning module is associated with gait cycle features, the joint control module with joint motion trajectory features, and the balance adjustment module with body posture balance features. The specific approach is to first clarify what information each functional module needs to function properly: for example, when planning stride frequency and stride length, the gait planning module must know the time pattern of the "support-swing" cycle during walking, the time difference between the left and right leg movements, etc., which are precisely the contents of the gait cycle features in the common feature library, thus matching them; when controlling joint movement, the joint control module needs to know the angle change curve that each joint should reach within a gait cycle, the angle values ​​at key time points, etc., which correspond to the joint motion trajectory features in the common feature library, thus matching them; when adjusting body posture, the balance adjustment module needs to know the tilt angle and tilt speed of the torso, etc., which reflect the body's balance state, which are precisely the contents of the body posture balance features, thus matching them. The principle behind this matching is that the features in the common feature library are general basic information extracted from various robot walking data. The operation of functional modules must rely on this basic information to realize their functions. Therefore, the type of common feature that is relied upon can be determined by the correspondence of the input information.

[0061] The gait planning module's period generation logic is transformed into conditional rule expressions based on gait period characteristics; the balance adjustment module's posture correction logic is transformed into functional relationships based on body posture balance characteristics; and the joint control module's trajectory tracking logic is transformed into parameterized equations based on joint motion trajectory characteristics, forming a symbolic rule set. Specifically, the internal working logic of each module is first analyzed: for example, the gait planning module's period generation logic might be "when the robot's legs are longer, the gait period is slightly longer; when the legs are shorter, the gait period is slightly shorter." During the transformation, the "leg length range" in the gait period characteristics is used as a condition, written as "if the leg length belongs to the long range, the gait period is within a certain range; if the leg length belongs to the short range..." The conditional rule expression is something like, "If the gait cycle falls within another range, then..."; the posture correction logic of the balance adjustment module might be, "The greater the body tilt angle, the greater the angle the joint needs to adjust." During the transformation, the "tilt angle" in the body posture balance characteristics is used as a variable, written as a functional relationship expression like "Joint adjustment angle = a certain amount × body tilt angle"; the trajectory tracking logic of the joint control module might be, "The hip joint angle changes sinusoidally over time, and the amplitude of the curve is related to leg length." During the transformation, "time" and "leg length" in the joint motion trajectory characteristics are used as variables, written as a parameterized equation like "Hip joint angle = amplitude × sin(2π × time / period)" (where amplitude is related to leg length). Collecting these expressions and equations forms a symbolic rule set. The principle behind this transformation is that the logic of each module is essentially based on the regular relationship of common characteristics. Symbolic representation can strip away specific numerical values, retaining only the logical structure and variable relationships, laying the foundation for subsequently distinguishing control strategies and control parameters.

[0062] Three sets of quadruped robot samples with different body characteristic parameters (leg length, weight, and motor torque) were selected, input into the controller, and the output changes of each module were recorded. Through comparative analysis, numerical variables in the symbolic rule set that change with the body characteristic parameters were identified and marked as candidate control parameters. Specifically, three sets of samples were selected, each with different leg lengths, weights, and motor torques, covering common quadruped robot body characteristic ranges. For example, the first set had short legs, light weight, and low torque; the second set had medium leg length, medium weight, and medium torque; and the third set had long legs, heavy weight, and high torque. The body characteristic parameters of these three sets of samples were then input into the controller. Observe and record the specific cycle values ​​output by the gait planning module, the specific joint angle values ​​output by the joint control module, and the specific correction magnitude output by the balance adjustment module. Then compare these three sets of output results to see which values ​​in the symbolic rule set change with the body characteristic parameters. For example, in the conditional rule expression of the gait cycle, when the leg length changes from short to long, a certain value changes from 1.2 to 1.5 and then to 1.8; this value is a variable that changes with the body characteristic. In the parameterized equation of the joint angle, when the motor torque changes from small to large, a certain coefficient changes from 0.3 to 0.5 and then to 0.7; this coefficient is also a variable that changes with the body characteristic. Mark all these variables as candidate control parameters. The principle behind this is that the essence of control parameters is that they are values ​​that adjust with changes in the robot's body characteristics. By changing the body characteristic parameters and observing which values ​​change accordingly, possible control parameters can be selected.

[0063] The control strategy core is defined as the logical structure in the symbolic rule set that does not change with the robot's body characteristic parameters. The control parameters are defined as the numerical variables among the candidate control parameters that have a monotonic mapping relationship with the robot's body characteristic parameters. This separation of control strategy and control parameters is achieved by first analyzing the symbolic rule set and identifying the parts that remain unchanged regardless of changes in the body characteristic parameters. For example, the conditional rule expression for the gait period is always a "condition-result" structure such as "if the leg length is within a certain interval, then the period takes a certain range," where only the specific values ​​of the interval and range change. This structure does not change with the body characteristic and is defined as the core of the control strategy. For candidate control parameters, their relationship with the body characteristic parameters is observed. For example, if a candidate parameter continuously increases with increasing leg length and continuously decreases with decreasing leg length, or continuously decreases with increasing weight and continuously increases with decreasing weight, this stable unidirectional change of one variable with another is a monotonic mapping relationship. Candidate parameters with this relationship are defined as control parameters. This separation effectively separates the control strategy (the invariant logical structure) from the control parameters (the numerical values ​​that change with the body). The core of the control strategy is a common logical framework for the walking control of different robots, which does not depend on the specific body; while the control parameters need to change regularly with the characteristics of the body in order to adapt to different robots. The monotonic mapping relationship ensures the stability and predictability of this change, and is therefore suitable as a control parameter.

[0064] In step S3, the process of establishing a control parameter mapping model by analyzing the correspondence between key ontological characteristic parameters and control parameters is as follows:

[0065] Quadruped robots with different leg lengths, weights, and motor torques were selected as samples to form a key body characteristic parameter sample set. From the existing quadruped robots, robots with leg lengths distributed within a common range were selected, such as some robots with short legs, some with medium legs, and some with long legs. At the same time, robots with different weights were selected, such as some with light weights, some with medium weights, and some with heavy weights. Robots with different motor torques were also selected, such as some with low torques, some with medium torques, and some with high torques. This ensures that these three parameters can cover different situations. The leg length, weight, and motor torque of these robots were recorded to form the key body characteristic parameter sample set.

[0066] For each quadruped robot in the sample set, all its control parameters under a standard walking scenario were acquired. A control parameter dataset was constructed, corresponding one-to-one with the key proprioceptive parameter sample set. Specifically, a standard walking scenario was first defined, such as walking in a straight line at a stable speed on a flat surface, eliminating interference from complex factors such as terrain undulations and turns. Then, each robot in the sample set was allowed to walk in this scenario. During the walking process, all control parameters used by each robot were recorded, such as the period coefficient in gait planning, the angle amplitude coefficient in joint control, and the threshold in balance adjustment. Afterward, each robot's key proprioceptive parameters (leg length, weight, motor torque) were mapped one-to-one with their corresponding control parameters. For example, if a robot's leg length, weight, and torque were a certain value, its corresponding control parameters would be another set of values. This formed the control parameter dataset. The standard walking scenario reduces the influence of environmental factors on the control parameters, making the differences in control parameters mainly determined by the robot's proprioceptive parameters. The one-to-one mapping clearly defines the control parameters corresponding to each combination of proprioceptive parameters, providing a data foundation for subsequent analysis of the relationship between the two.

[0067] Using key body characteristic parameters as row vectors and control parameters as column vectors, the Pearson correlation coefficient between any two parameters is calculated to form a parameter correlation matrix. The matrix element values ​​represent the degree of linear correlation between the corresponding two parameters. The key body characteristic parameters (leg length, weight, motor torque) are used as rows of the matrix, with each row representing one body characteristic parameter; all control parameters are used as columns of the matrix, with each column representing one control parameter. Then, the Pearson correlation coefficient between the body characteristic parameter corresponding to each row and the control parameter corresponding to each column is calculated, and these coefficients are filled into the corresponding positions in the matrix to form the parameter correlation matrix.

[0068] Based on the parameter correlation matrix, a hierarchical clustering algorithm is used to group the control parameters. Control parameters with high correlation coefficients to the same key ontological characteristic parameter are grouped into a parameter influence cluster. Each cluster corresponds to the influence range of a key ontological characteristic parameter. Specifically, the correlation coefficients between each control parameter and the key ontological characteristic parameter in the parameter correlation matrix are first observed. For example, some control parameters have a high correlation coefficient with leg length, but a low correlation coefficient with weight and motor torque. Then, the hierarchical clustering algorithm is used, starting with each control parameter as a separate group, and gradually merging control parameters with high correlation coefficients to the same key ontological characteristic parameter into the same group. For example, control parameters with high correlation coefficients to leg length are merged into one group, those with high correlation coefficients to weight are merged into another group, and those with high correlation coefficients to motor torque are merged into a third group. These groups are called parameter influence clusters. The hierarchical clustering algorithm can naturally form groups based on the similarity between parameters (here, similarity is reflected in the degree of correlation with the same ontological characteristic parameter). Control parameters with high correlation coefficients to the same key ontological characteristic parameter indicate that their changes are mainly affected by that ontological characteristic parameter, and therefore they are grouped into the same cluster, which facilitates the subsequent establishment of mapping relationships for each cluster individually.

[0069] For each parameter influence cluster, a parameter mapping sub-model is trained using the corresponding key ontological characteristic parameters as input and the control parameters within the cluster as output. Specifically, for each parameter influence cluster, its corresponding key ontological characteristic parameters are determined. For example, if a cluster corresponds to leg length, then the leg lengths of different robots in the sample set are used as input data. At the same time, all control parameters within the cluster are used as output data. For example, if the cluster includes control parameters such as gait period coefficient and joint angle amplitude, then the values ​​of these parameters are used as output. Then, the gradient boosting regression algorithm is used to iteratively adjust the model, allowing the model to learn the relationship between the input leg length and the output cluster control parameters. After training, a parameter mapping sub-model for that cluster is obtained. The same method can be used to train parameter mapping sub-models for corresponding weight and motor torque. The control parameters within each cluster are mainly influenced by the corresponding key ontological characteristic parameters. The gradient boosting regression algorithm can gradually reduce the prediction error by continuously superimposing the prediction results of multiple simple models, thereby accurately capturing the complex nonlinear relationship that may exist between the input (ontological characteristic parameters) and the output (control parameters), enabling the sub-model to accurately predict the corresponding control parameters based on the input ontological characteristic parameters.

[0070] To construct a control parameter mapping model, all parameter mapping sub-models are concatenated and integrated. Specifically, the trained sub-models are connected according to their corresponding key ontological characteristic parameters to form a unified model. For example, when generating control parameters for a new quadruped robot, the robot's leg length, weight, and motor torque are first obtained. Then, the leg length is input into the corresponding parameter mapping sub-model to obtain the control parameters for that sub-model; the weight is input into its corresponding sub-model to obtain the corresponding control parameters; the motor torque is input into its corresponding sub-model to obtain the corresponding control parameters; finally, the control parameters output by these sub-models are summed up to obtain all the control parameters required for the new robot. This concatenated and integrated model is the control parameter mapping model.

[0071] In step S4, the method for constructing the control policy transfer model is as follows:

[0072] A strategy feature encoding module is constructed to convert gait cycle features, joint motion trajectory features, and body posture balance features from the common feature library into a fixed-dimensional feature matrix. First, the specific information contained in each common feature is identified. For example, gait cycle features may include cadence, single-leg support time percentage, and left-right leg phase difference; joint motion trajectory features may include maximum hip joint angle and knee joint velocity change rate; and body posture balance features may include maximum trunk pitch angle and balance adjustment response time. Then, a fixed number of dimensions are assigned to each feature. For example, five dimensions are used to describe gait cycle features, five dimensions to describe joint motion trajectory features, and five dimensions to describe body posture balance features. The specific information of these features is then sequentially filled into the matrix according to their dimensions, ensuring that the final matrix has a fixed number of rows and columns regardless of changes in the details of the input features. The principle behind this is that a fixed-dimensional feature matrix allows the model to stably receive input, avoiding processing chaos caused by changes in the number or structure of features. Simultaneously, integrating different types of common features into a single matrix facilitates the model's unified extraction of cross-feature correlation information. A strategy logic transformation layer is constructed using a gated loop unit structure. The input is a symbolic sequence of the separated control strategy logic expressions. The symbolic rules of the control strategy are decomposed into a series of consecutive symbolic steps according to the execution order. For example, "If the body tilt angle exceeds the threshold, adjust the hind leg joint angle" can be decomposed into a symbolic sequence like "[body tilt angle] → [exceeds the threshold] → [adjust hind leg] → [joint angle]". The gated loop unit structure receives these symbols sequentially. When processing each symbol, it combines previously received symbol information to determine which information needs to be retained and which can be ignored. For example, when processing "adjust hind leg", it associates it with the previous "exceeds the threshold" information to ensure logical coherence. The logical expression of the control strategy is essentially a sequence executed in time or conditional order. The gated loop unit excels at processing sequential data and can capture long-distance dependencies in the sequence, thus accurately understanding the sequential relationships of the strategy logic. A body characteristic modulation layer is set up to receive leg length, weight, and motor torque parameters and generate modulation coefficients. Specifically, these three parameters are input into this layer, and the layer generates corresponding coefficients based on their values. For example, when the legs are long, a larger coefficient is generated to adjust the gait cycle-related strategy logic; when the weight is large, a coefficient is generated to enhance the strength of the balance adjustment strategy. These modulation coefficients are applied to the intermediate results output by the strategy logic conversion layer, adjusting the logical weights of different parts. The principle behind this is that robots with different body characteristics require strategies that adapt to specific details. The modulation coefficients are equivalent to "weighting" different parts of the strategy based on the body characteristics, allowing the strategy logic to adapt to the specific characteristics of the robot while maintaining the core structure.A strategy output decoding layer is constructed to convert processed signals into a set of executable control strategy instructions. Specifically, it receives signals adjusted by the ontology feature modulation layer. These signals are abstract information that has been encoded and converted. The decoding layer restores them to specific instructions that the robot can directly execute. For example, it converts the abstract signal "adjust joint angle" into a specific instruction like "rotate the hip joint to 30 degrees within 0.5 seconds," while ensuring that the format and order of the instructions meet the robot's execution requirements. The modules are connected sequentially according to the data flow: common features are first processed by the encoding module, then input into the strategy logic conversion layer, then combined with coefficients generated by the ontology feature modulation layer, and finally the decoding layer outputs the instructions, forming a complete architecture. Data is transferred according to the flow of "feature input → logic understanding → ontology adaptation → instruction output." Each module processes the results of the previous stage and provides input for the next stage, ensuring a continuous and complete conversion process from common features to final instructions.

[0073] Walking data of various quadruped robots in flat ground, slope, and staircase scenarios were collected. Common feature matrices, ontological parameters, and the actual control strategy command set executed were extracted for each data point. Specifically, robots with different leg lengths, weights, and motor torques were allowed to walk in the three scenarios, and motion data during walking was recorded, including the time of each step, joint rotation angles, and body tilt. From this data, common feature matrices (such as the previously encoded fixed-dimensional matrix), ontological parameters such as the robot's leg length, and the actual control commands executed by the robot (such as "gait cycle 1 second" and "knee joint rotation angle 20 degrees") were extracted. The dataset was divided into training and validation datasets in a 7:3 ratio to ensure that both datasets contained a similar proportion of different robot samples. Specifically, the number of different robot samples was first counted. For example, if there were 10 types of robots, each representing 10%, then each type of robot would also represent 10% of the training set, and the same applied to the validation set. 70% of the data was randomly selected as training data for model learning, and the remaining 30% was used as validation data to test the model's learning performance. The principle behind this approach is that data from multiple scenarios and robots ensures that the model learns general rules, the 7:3 ratio ensures that there is enough training data to support model learning, validation data is used to monitor whether the model relies too much on training data and loses its generality, and the same proportion of robot samples avoids the fact that a certain type of robot has too high or too low a proportion in the validation set, which would lead to inaccurate validation results.

[0074] A strategy logic deviation loss term is constructed, calculating the edit distance between the control strategy instruction set output by the model and the actual instruction set. By comparing the instruction set output by the model with the instruction set actually executed by the robot, the minimum number of operations required to modify the model instructions to the actual instructions is calculated, including adding, deleting, or replacing instructions. For example, if the model outputs "gait cycle 1.2 seconds" while the actual value is "1.0 seconds," one replacement operation is required, and the edit distance is 1. The more operations required, the greater the deviation between the model's output strategy logic and the actual value. A common feature consistency loss term is constructed, using cosine similarity to measure the difference between the common feature matrix before and after model processing. Specifically, the original common feature matrix input to the model is compared with the feature matrix after model processing (e.g., after encoding, transformation, etc.). The closer the cosine similarity is to 1, the more consistent the directions of the two matrices, meaning that the core information of the common features has been preserved during model processing; the lower the similarity, the more severely the common features have been distorted. The two losses are added together with a 1:1 weight to form a joint loss function. The principle behind this is that the policy logic deviation loss ensures that the instructions output by the model conform to the actual execution logic, while the common feature consistency loss ensures that the model does not lose or distort common features during processing (these are the basis of policy transfer). Both are equally important, and adding them together allows the model to optimize the performance of both aspects simultaneously, ensuring both the correctness of the instructions and the stability of the common features.

[0075] The first stage trains the model using flat ground scene data from the training dataset. Gradient descent is used to minimize the joint loss function, with a preset base number of iterations. Specifically, the model first learns the simplest walking scenario. During each training session, flat ground data is input into the model, the joint loss between the model output and the actual data is calculated, and then the model parameters are adjusted using gradient descent to gradually reduce the loss. This process is repeated for the preset base number of iterations, allowing the model to grasp basic policy transfer rules. The second stage introduces slope and staircase scene data, maintaining the same learning rate. The number of iterations is half the base number. This is because after the model has mastered the transfer rules on flat ground, more complex scene data is introduced. At this point, the model has a certain foundation and can adapt to complex scenes more quickly, hence the halving of the iteration count. Maintaining the same learning rate ensures that the parameter adjustments are consistent with the first stage when adapting to new scenes, avoiding adaptation difficulties caused by changes in the learning rate. After each iteration, the loss value is calculated using the validation dataset, and the model parameters are recorded. Specifically, after each iteration, the model is input with validation data that was not used in training, and the loss value is calculated. If the loss value is smaller than before, it means that the current model parameters are better, and this is recorded. The principle behind this is that phased training from simple to complex conforms to the law of gradual model learning. Mastering the basics first and then challenging complex scenarios allows the model to accumulate knowledge more stably. The validation process can promptly detect whether the model is overfitting during training (i.e., only adapting to training data and not to new data). By recording the optimal parameters, the generality of the model is ensured.

[0076] The model parameters with the minimum loss value in the validation dataset are selected as initial parameters. Specifically, the set of parameters with the minimum loss value on the validation dataset is identified from the recorded model parameters. The model corresponding to this set of parameters performs best on unseen data, and fine-tuning is performed starting from this set. A separate round of fine-tuning training is conducted for each robot sample. The ontological characteristic parameters and common feature matrix of the robot are input, and the deviation from the actual control strategy instruction set is used as the fine-tuning loss. The ontological characteristic modulation layer parameters are updated through a single iteration. Specifically, for each robot, the model is trained using its specific data, and only the parameters of the ontological characteristic modulation layer are adjusted (because this layer is responsible for adapting to the specific ontology). The deviation between the model output and the actual instructions is calculated as the fine-tuning loss, and the parameters are adjusted through a single iteration to reduce the loss. The principle behind this approach is that the parameters with the minimum loss on the validation set already possess good generality. Fine-tuning only on the ontological characteristic modulation layer allows the model to maintain a general strategy while more accurately adapting to the characteristics of each robot. A single iteration avoids over-fine-tuning that would destroy the general rules already learned by the model, ensuring that the core strategy remains unchanged. The model training is completed through these steps.

[0077] In step S5, the process of forming a walking controller for the new quadruped robot is as follows:

[0078] The control strategy, adapted through the control strategy transfer model, is analyzed. All logical nodes requiring control parameters are extracted, and the parameter types for each node are recorded. The adapted control strategy is carefully examined to identify all parts that require specific values ​​to take effect; these are the logical nodes. For example, in the strategy "Initiate balance adjustment when the body tilt angle exceeds X," "X" is a logical node that requires a specific angle value. The parameter types required by each node are also recorded; for instance, "X" requires an angle-type parameter, while another node, "Stride length Y," requires a length-type parameter. The principle behind this is that while the core logic of the control strategy is general, its execution depends on specific values. These nodes requiring values ​​are crucial for connecting control parameters and strategy logic; clearly defining their location and type ensures accurate parameter input. The initial values ​​of control parameters generated by the control parameter mapping model are categorized by type, and a parameter strategy association index table is constructed to clarify the logical node position corresponding to each initial value of control parameter. Specifically, the parameters generated by the mapping model are organized by type; for example, all angle-type parameters are grouped together, and length-type parameters are grouped together. Then, based on the previously recorded logical node types, each initial value of parameter is associated with its corresponding logical node. For example, the angle parameter "2 degrees" corresponds to the "X" node in "body tilt angle exceeds X", and the length parameter "0.3 meters" corresponds to the "Y" node in "stride length is Y". This correspondence is recorded in tabular form, which is the parameter strategy association index table. The principle behind this is that the initial values ​​of control parameters are generated by type, and logical nodes also have clear type requirements. Establishing associations through type matching ensures that each parameter accurately corresponds to the node that needs it, avoiding parameter mismatch.

[0079] Design an interface adaptation layer that takes as input a set of initial control parameter values ​​and an adapted control strategy, and outputs a strategy execution sequence with embedded parameters. The interface adaptation layer needs to match the data types of the initial control parameter values ​​with those of the logical nodes, converting numerical parameters into symbolic parameter formats recognizable by the strategy. Specifically, the interface adaptation layer receives two parts: all initial control parameter values ​​and the adapted control strategy. The layer first checks whether the type of each parameter matches the type required by the corresponding logical node; for example, it ensures that angle parameters are not filled into length nodes. Then, it converts simple numerical values ​​(such as "2") into a form that the strategy can understand (such as "tilt angle threshold = 2 degrees"), enabling the parameters to be recognized and used by the strategy logic. The final output is a complete sequence of execution steps with parameters embedded in the corresponding nodes. Since the initial control parameter values ​​are simple numerical values, while the control strategy requires parameter expressions with clear meaning, the interface adaptation layer, as an intermediate conversion link, eliminates the difference in expression between the two, ensuring that the parameters can be correctly interpreted by the strategy.

[0080] Based on the parameter strategy association index table, the initial values ​​of control parameters are injected one by one into the corresponding logical nodes of the control strategy through the interface adaptation layer. During the injection process, the logical structure of the control strategy remains unchanged; only the parameter placeholders within the nodes are replaced to form a complete strategy expression with specific parameters. Following the guidance of the index table, the logical node corresponding to each initial value of control parameter is found, and the parameter is filled into the node through the interface adaptation layer. For example, if the index table finds the parameter "2 degrees" corresponding to the node "X", then "X" is replaced with "2 degrees". In this process, the original logical order and conditional relationships in the strategy (such as the "if...then..." structure) remain unchanged; only the places where parameters are needed are modified. The principle behind this is that the core logic of the control strategy is common knowledge extracted from various robots, and its structure is universal. It does not require modification; only specific parameters need to be replaced to adapt to new robots. Maintaining the unchanged structure ensures the stability of the core functionality of the strategy.

[0081] The complete strategy expression is input into the simulation environment, which includes a virtual model of a quadruped robot with identical physical characteristics to the new quadruped robot. The simulation environment is run to monitor for anomalies such as parameter out-of-bounds errors or logical conflicts during the execution of the strategy expression. The location of the logical nodes where anomalies occur is recorded. Specifically, a strategy expression with specific parameters is input into the simulation software, which contains a virtual robot with the exact same leg length, weight, and motor torque as the new robot. The simulation is started, and the virtual robot is made to perform walking actions according to the strategy expression. Anomalies are observed, such as a joint rotation angle exceeding its physical limit (parameter out-of-bounds error) or contradictory actions required by two strategy instructions (logical conflict). If an anomaly occurs, the parameter of the corresponding logical node is recorded; for example, a node specifying a stride length of 0.5 meters might prevent the legs from reaching that length. The simulation environment can safely simulate the new robot's walking process, identifying problems in the strategy early and avoiding potential damage from direct testing on a physical robot. The consistency of parameters between the virtual model and the new robot ensures that the simulation results match reality.

[0082] For anomalies encountered during verification, the numerical range of the initial values ​​of the corresponding control parameters is adjusted through the interface adaptation layer, and the parameter injection operation is re-executed until no anomalies appear in the simulation environment. The complete policy expression that passes verification is converted into code format executable by the controller, forming a walking controller for the new quadruped robot. Specifically, based on the anomaly nodes recorded in the simulation, the initial values ​​of the parameters corresponding to those nodes are modified through the interface adaptation layer. For example, the "0.5-meter stride" that causes joint overshoot is adjusted to "0.4 meters," and then the parameters are re-injected into the policy expression, and simulation testing is performed again. This process is repeated until the virtual robot can walk normally without any anomalies. Afterward, the policy expression that passes verification is converted into a form that the new robot controller can directly execute. For example, "stride 0.4 meters" is converted into an instruction that the controller can recognize. The final result is the walking controller. The initial parameter values ​​may have deviations due to errors in the mapping model. Adjustments are made to adapt to the actual characteristics of the new robot. Converting them into a format executable by the controller is necessary because the controller needs specific instruction forms to drive robot movement, ensuring that the policy can be truly implemented.

[0083] Please see Figure 2 As shown, the present invention also provides a quadruped robot walking controller training system, implemented in the aforementioned quadruped robot walking controller training method, comprising:

[0084] Common feature acquisition and database construction module: Collects motion data of various quadruped robots, extracts common features of gait cycle and joint motion trajectory, and stores them in the common feature database;

[0085] Controller parameter and strategy separation module: Based on the common feature library, the adaptive walking controller of a specific quadruped robot is analyzed and the control parameters and control strategy are separated.

[0086] Mapping model construction module: Based on the differences in the body characteristics of different quadruped robots, key parameters are selected, their correspondence with control parameters is analyzed, and a mapping model is established.

[0087] Control strategy transfer model construction module: Based on the common features in the common feature library and the separated control strategies, a control strategy transfer model is constructed and trained using walking data from various quadruped robots;

[0088] Walking controller integration module: Integrates the initial values ​​of the new quadruped robot control parameters generated by the control parameter mapping model with the adapted control strategy to form a walking controller.

[0089] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A training method for a quadruped robot walking controller, characterized in that, Includes the following steps: S1: Collect motion data of various quadruped robots in various walking scenarios, extract common features related to walking control, including gait cycle features, joint motion trajectory features, and body posture balance features, and store the common features in a common feature library; S2: Based on the common feature library, the adaptive walking controller trained for a specific quadruped robot is analyzed, and the control parameters and control strategies in the controller are separated. S3: Based on the differences in the body characteristics of different quadruped robots, key body characteristic parameters are selected, including leg length, weight, and motor torque. By analyzing the correspondence between key body characteristic parameters and control parameters, a control parameter mapping model is established. S4: Based on the common features in the common feature library and the separated control strategies, construct a control strategy transfer model, and train the control strategy transfer model using walking data from various quadruped robots. The method for constructing the control strategy migration model is as follows: A strategy feature encoding module is constructed to convert gait cycle features, joint motion trajectory features, and body posture balance features from the common feature library into a fixed-dimensional feature matrix; a strategy logic conversion layer is built, adopting a gated loop unit structure, with the input being a symbolic sequence of the separated control strategy logic expressions; a body characteristic modulation layer is set up to receive leg length, weight, and motor torque parameters and generate modulation coefficients; a strategy output decoding layer is constructed to convert the processed signals into an executable control strategy instruction set. The modules are connected sequentially according to the data flow to form a complete control strategy transfer model architecture. The process of training the control strategy transfer model using walking data from various quadruped robots is as follows: When preparing the training dataset, walking data of various quadruped robots in flat ground, slope, and staircase scenarios were collected. Common feature matrices, ontological characteristic parameters, and the actual control strategy instruction set executed were extracted for each data point. A joint loss function was designed, constructing a strategy logic deviation loss term using edit distance and a common feature consistency loss term using cosine similarity, and then adding them with a 1:1 weight to form the joint loss function. Phased training was performed: first, the flat ground scenario data from the training dataset was used for training with gradient descent algorithm, with the number of iterations being a preset base number. Then, slope and staircase scenario data were added for further training, with the number of iterations being half the base number. Each iteration used the validation dataset to calculate the loss value and record the model parameters. Model adaptation fine-tuning was performed, selecting the model parameters with the smallest loss value from the validation dataset as initial parameters. Fine-tuning was performed individually for each robot sample, inputting its ontological characteristic parameters and common feature matrix, using the deviation between the output and the actual control strategy instruction set as the fine-tuning loss. The ontological characteristic modulation layer parameters were updated through a single iteration to complete model training. S5: The initial values ​​of the control parameters of the new quadruped robot generated by the control parameter mapping model are integrated with the control strategy adapted by the control strategy transfer model to form a walking controller for the new quadruped robot.

2. The training method for a quadruped robot walking controller according to claim 1, characterized in that, In step S2, the control parameters are specific values ​​related to the robot's body characteristic parameters, and the control strategy is the logic and rules for walking control based on common features.

3. The training method for a quadruped robot walking controller according to claim 1, characterized in that, In step S2, the process of separating the control parameters and control strategy in the controller is as follows: The existing adaptive walking controller is functionally decomposed into a gait planning module, a joint control module, and a balance adjustment module. The inputs of each functional module are matched with common features in the common feature library to determine the types of common features that the module depends on for operation; among them, the gait planning module is associated with gait cycle features, the joint control module is associated with joint motion trajectory features, and the balance adjustment module is associated with body posture balance features. The gait planning module's periodic generation logic is transformed into a conditional rule expression based on gait periodic characteristics; the balance adjustment module's posture correction logic is transformed into a functional relationship expression based on body posture balance characteristics; and the joint control module's trajectory tracking logic is transformed into a parameterized equation based on joint motion trajectory characteristics, forming a symbolic rule set. Three sets of quadruped robot samples with different body characteristic parameters were selected, input into the controller, and the output changes of each module were recorded. Through comparative analysis, the numerical variables in the symbolic rule set that change with the body characteristic parameters were identified and marked as candidate control parameters. The logical structure in the symbolic rule set that does not change with the robot's ontological characteristic parameters is defined as the core of the control strategy; the numerical variables in the candidate control parameters that have a monotonic mapping relationship with the robot's ontological characteristic parameters are defined as control parameters, thus completing the separation of control strategy and control parameters.

4. The training method for a quadruped robot walking controller according to claim 1, characterized in that, The process of establishing a control parameter mapping model by analyzing the correspondence between key ontological characteristic parameters and control parameters is as follows: Quadruped robots with different leg lengths, weights, and motor torques were selected as samples to form a sample set of key body characteristic parameters; For each quadruped robot in the sample set, all its control parameters under the standard walking scenario are obtained, and a control parameter dataset corresponding one-to-one with the key body characteristic parameter sample set is constructed. Using key ontological characteristic parameters as row vectors and control parameters as column vectors, the Pearson correlation coefficient between any two parameters is calculated to form a parameter correlation matrix. The matrix element values ​​represent the degree of linear correlation between the corresponding two parameters. Based on the parameter correlation matrix, a hierarchical clustering algorithm is used to group the control parameters. Control parameters with high correlation coefficients with the same key ontological characteristic parameter are grouped into a parameter influence cluster, and each cluster corresponds to the influence range of a key ontological characteristic parameter. For each parameter-affected cluster, the corresponding key ontological characteristic parameters are used as inputs, and the control parameters within the cluster are used as outputs. The gradient boosting regression algorithm is used to train the parameter mapping sub-model. All parameter mapping sub-models are chained together and integrated to construct a control parameter mapping model.

5. The training method for a quadruped robot walking controller according to claim 1, characterized in that, The process of forming the walking controller for this new quadruped robot is as follows: The control strategy after adaptation by the control strategy transfer model is analyzed, all logical nodes that require input control parameters are extracted, and the parameter types corresponding to each node are recorded; the initial values ​​of control parameters generated by the control parameter mapping model are classified by type, and a parameter strategy association index table is constructed. Design an interface adaptation layer whose input is a set of initial control parameter values ​​and an adapted control strategy, and whose output is the strategy execution sequence after parameter embedding. Based on the parameter strategy association index table, the initial values ​​of control parameters are injected one by one into the corresponding logical nodes of the control strategy through the interface adaptation layer. During the injection process, the logical structure of the control strategy remains unchanged, and only the parameter placeholders in the nodes are replaced to form a complete strategy expression with specific parameters. Input the complete strategy expression into the simulation environment. The simulation environment has a built-in virtual model of a quadruped robot, whose body characteristic parameters are consistent with the new quadruped robot. Run the simulation environment and monitor whether parameter out-of-bounds or logical conflict anomalies occur during the execution of the strategy expression. Record the location of the logical node where the anomaly occurs. For any anomalies encountered during verification, the numerical range of the initial values ​​of the corresponding control parameters is adjusted through the interface adaptation layer, and the parameter injection operation is re-executed until no anomalies occur in the simulation environment. The complete policy expression that has passed verification is converted into code format that the controller can execute, forming a walking controller for the new quadruped robot.

6. A quadruped robot walking controller training system, implemented according to any one of claims 1-5, characterized in that, include: Common feature acquisition and database construction module: Collects motion data of various quadruped robots, extracts common features of gait cycle and joint motion trajectory, and stores them in the common feature database; Controller parameter and strategy separation module: Based on the common feature library, the adaptive walking controller of a specific quadruped robot is analyzed and the control parameters and control strategy are separated. Mapping model construction module: Based on the differences in the body characteristics of different quadruped robots, key parameters are selected, their correspondence with control parameters is analyzed, and a mapping model is established. Control strategy transfer model construction module: Based on the common features in the common feature library and the separated control strategies, a control strategy transfer model is constructed and trained using walking data from various quadruped robots; Walking controller integration module: Integrates the initial values ​​of the new quadruped robot control parameters generated by the control parameter mapping model with the adapted control strategy to form a walking controller.

Images