A bionic robot control method and system based on AI learning

Abstract

The present application provides a bionic robot control method and system based on AI learning, including: acquiring dynamic environmental data including obstacle positions, target states and terrain features, and collecting and transmitting real-time data to a data processing module through a sensor network; for the dynamic environmental data, removing noise and redundant information, extracting key features, and generating a structured environmental feature data set; based on the structured environmental feature data set, using a Gaussian mixture model to construct an environmental model, and if the environmental change exceeds a preset threshold, updating the parameters of the environmental model to reflect the latest environmental status; based on the environmental model, using a time series prediction algorithm to predict future environmental change trends, and if the deviation between the prediction result and the current environmental status exceeds a preset range, triggering a model retraining mechanism; based on the prediction results, combined with the current state and target task of the bionic robot, using a reinforcement learning algorithm to generate an optimal decision-making strategy.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a bionic robot control method and system based on AI learning. Background Art

[0002] When making predictions and decisions in dynamic environments, biomimetic robot control systems face a conflict between prediction accuracy and real-time performance. Improving prediction accuracy requires collecting and processing large amounts of environmental data, and employing complex machine learning algorithms for modeling and prediction. This often requires long computation times and significant resource overhead, hindering the system's real-time responsiveness. Furthermore, biomimetic robots must react and make decisions quickly in dynamic environments, placing high demands on the real-time performance of the control system. If the prediction process takes too long, the robot will be unable to make accurate decisions and control actions in a timely manner, impacting its adaptability and task completion efficiency in complex environments. In practical applications, biomimetic robot control systems also face challenges posed by environmental uncertainty and variability. Environmental factors such as obstacles, targets, and terrain can change at any time, and these changes are often difficult to accurately predict. Even with advanced machine learning algorithms, it is difficult to fully cover all possible scenarios. This requires robustness and adaptability in the control system, enabling timely adjustment of strategies in the event of deviations in predictions or sudden changes in the environment to ensure stable robot operation. Summary of the Invention

[0003] The present invention provides a bionic robot control method based on AI learning, which mainly includes:

[0004] Acquire dynamic environmental data including obstacle location, target status and terrain characteristics, and collect and transmit real-time data to the data processing module through the sensor network;

[0005] For dynamic environmental data, remove noise and redundant information, extract key features, and generate a structured environmental feature dataset;

[0006] Based on a structured environmental feature dataset, a Gaussian mixture model is used to build an environmental model. If the environmental change exceeds a preset threshold, the parameters of the environmental model are updated to reflect the latest environmental status.

[0007] Based on the environmental model, a time series prediction algorithm is used to predict future environmental change trends. If the deviation between the predicted result and the current environmental state exceeds a preset range, the model retraining mechanism is triggered;

[0008] Based on the prediction results, the reinforcement learning algorithm is used to generate the optimal decision strategy in combination with the current state and target task of the bionic robot;

[0009] Control instructions are generated through decision-making strategies. If the control instructions conflict with the current execution status, the robustness control module is called to adjust the instruction priority. The robustness control module adapts to environmental changes by dynamically adjusting control parameters to ensure the stability and adaptability of instruction execution.

[0010] The present invention provides a bionic robot control system based on AI learning, which mainly includes:

[0011] Environmental data acquisition module, used to obtain dynamic environmental data including obstacle location, target status and terrain characteristics, and collect and transmit real-time data to the data processing module through the sensor network;

[0012] The data processing module is used to remove noise and redundant information from dynamic environmental data, extract key features, and generate a structured environmental feature data set;

[0013] The environmental modeling module is used to build an environmental model based on a structured environmental feature dataset using a Gaussian mixture model. If the environmental change exceeds a preset threshold, the parameters of the environmental model are updated to reflect the latest environmental status.

[0014] The prediction analysis module is used to predict future environmental change trends based on the environmental model using a time series prediction algorithm. If the deviation between the prediction result and the current environmental state exceeds a preset range, the model retraining mechanism is triggered;

[0015] The decision generation module is used to generate the optimal decision strategy based on the prediction results, combined with the current state of the bionic robot and the target task, using the reinforcement learning algorithm;

[0016] The robustness control module is used to generate control instructions through decision-making strategies. If the control instructions conflict with the current execution status, the robustness control module is called to adjust the instruction priority. The robustness control module adapts to environmental changes by dynamically adjusting control parameters to ensure the stability and adaptability of instruction execution.

[0017] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0018] The present invention discloses an adaptive control method for a bionic robot based on AI learning. This method collects environmental data in real time through a sensor network, performs denoising and feature extraction on the data, and constructs a Gaussian mixture environment model. The present invention utilizes a time series prediction algorithm to predict environmental change trends and dynamically updates the model based on the prediction results. Based on the current state and mission objectives of the bionic robot, the present invention employs a reinforcement learning algorithm to generate an optimal decision-making strategy and adjusts the priority of control instructions through a robust control module to ensure execution stability and adaptability. This method can effectively cope with complex and changing environments and improve the environmental adaptability and mission execution efficiency of the bionic robot. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of a bionic robot control method based on AI learning of the present invention.

[0020] Figure 2 Schematic diagram of a bionic robot control method and system based on AI learning of the present invention.

[0021] Figure 3 This is another schematic diagram of a bionic robot control method and system based on AI learning of the present invention.

[0022] Figure 4 This is a structural diagram of a bionic robot control method and system based on AI learning of the present invention. DETAILED DESCRIPTION

[0023] To help those skilled in the art better understand the technical solutions in this specification, the following will provide a clear and complete description of the technical solutions in the embodiments of this specification, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. All other embodiments derived by those skilled in the art based on the embodiments in this specification without creative effort shall fall within the scope of protection of this specification.

[0024] like Figure 1-4 In this embodiment, a bionic robot control method based on AI learning may specifically include:

[0025] Step S101 , acquiring dynamic environment data including obstacle positions, target states and terrain features, and collecting and transmitting real-time data to a data processing module through a sensor network.

[0026] By deploying sensors such as lidar and cameras on a biomimetic robot, dynamic environmental data, including obstacle locations, target status, and terrain features, is collected in real time. The robot's wireless communication module transmits the raw data collected by the sensors to an onboard data processing unit in real time. The onboard data processing unit first preprocesses the received lidar point cloud data and camera image data. Statistical filtering is used to remove noise and outliers from the lidar point cloud data, and downsampling is performed to reduce the data volume. Median filtering is used to remove noise from the camera image data, and histogram equalization is performed to enhance image contrast. The preprocessed lidar point cloud data is then fed into the SLAM module, which uses the Iterative Closest Point (ICP) algorithm to locate the robot in real time and construct an environmental map. The preprocessed camera image data is then fed into the object detection module, which uses the YOLO (You-Only LookOnce) algorithm to detect and track target objects in real time. Based on the environmental map constructed by the SLAM module, the lidar point cloud data is rasterized to generate a two-dimensional occupancy grid map, representing the traversable and obstacle areas in the environment. At the same time, the target object's bounding box information output by the object detection module is projected onto the occupancy grid map, updating the target object's position. Next, a support vector machine (SVM) algorithm is used to classify each grid in the occupancy grid map, determining its traversability probability and extracting key terrain features such as slope and roughness. Based on the traversability probabilities and terrain features, the occupancy grid map is further converted into a cost map, representing the traversability cost of each grid. Finally, an A* search algorithm is used to search for an optimal path on the grid map, using the robot's current position as the starting point and the target position as the destination. The occupancy grid map serves as the environmental model, and the cost map as the heuristic function. The A* algorithm represents nodes in grid coordinates, and the node cost consists of two parts: the actual cost from the current node to the starting point and the estimated cost from the current node to the destination, calculated from the cost map. During the search, nodes with the lowest cost are preferentially expanded until the target node is found. Through these steps, the robot can obtain an optimal navigation path from its current position to the target location and autonomously navigate according to the path, while simultaneously updating the environmental map and avoiding obstacles in real time.

[0027] For example, biomimetic robots deploy sensors such as lidar and cameras to collect dynamic environmental data in real time. For example, a city rescue robot searches for survivors in a rubble field. The lidar scans the surrounding environment, generates point cloud data, and accurately locates obstacles. The camera captures high-definition images for identifying possible signs of life. The robot transmits the collected raw data via a wireless communication module to an onboard data processing unit. In practical applications, this might be an embedded system such as a Raspberry Pi or an industrial-grade computer. The data processing unit first preprocesses the received data. Statistical filtering is used to remove noise from the lidar point cloud data, such as outliers that significantly deviate from the main point cluster. Median filtering is used to remove noise from the camera images, improving image quality. The preprocessed lidar data is then fed into the SLAM module, which uses the ICP algorithm to locate the robot's position in real time and construct a map of the environment. In rescue scenarios, this enables the robot to accurately locate itself in complex rubble environments and create a three-dimensional map of its surroundings. Simultaneously, the preprocessed image data is fed into the object detection module, which uses the YOLO algorithm to detect and track objects. This enables the rescue robot to quickly identify possible survivors or important items. Based on the environmental map constructed by SLAM, the LiDAR point cloud data is rasterized to generate a two-dimensional occupancy grid map. In rescue missions, this map clearly identifies traversable and obstacle-prone areas, helping the robot plan a safe path. Target object information output by the object detection module is projected onto the occupancy grid map, updating the target position. Next, a support vector machine (SVM) algorithm is used to classify each grid in the occupancy grid map, determining its traversability probability. Building deformation is assessed to assess its integrity and damage level, determining whether it is at risk of collapse. Buildings with deformation exceeding a preset threshold are marked as potentially hazardous. In a rescue context, this step can identify potentially hazardous areas, such as unstable structures or areas at risk of collapse. Key terrain features, such as slope and roughness, are also extracted; these features are crucial for determining whether the robot can traverse certain areas. Based on the traversability probability and terrain characteristics, the occupancy grid map is converted into a costmap. In rescue scenarios, this costmap not only accounts for physical obstacles but also factors such as terrain difficulty and potential risk. For example, a steeply sloped area, while physically navigable, may be assigned a higher cost in the costmap. Finally, the A* search algorithm is used to search for the optimal path on the grid map, starting from the robot's current position and ending at the target position. In a rescue mission, this might be the path from the current position to the location of a detected survivor. The A* algorithm considers both actual and estimated costs, enabling it to quickly find a path that balances safety and efficiency in complex environments.Through this series of steps, the rescue robot can autonomously navigate in complex ruins environments, update environmental information in real time, effectively avoid obstacles, and quickly reach the target location, greatly improving the search and rescue efficiency and success rate.

[0028] Step S102 : For the dynamic environment data, remove noise and redundant information, extract key features, and generate a structured environment feature data set.

[0029] Based on the dynamic environmental data acquired by the biomimetic robot, the Kalman filter algorithm is used to filter the data, removing noise and redundant information to obtain clean environmental data. The SIFT algorithm is used to extract key feature points from the filtered environmental data, obtaining feature descriptors that are invariant to scale, rotation, and brightness. The extracted SIFT feature points are clustered using the K-means clustering algorithm to obtain sets of feature points of different categories, each representing an environmental feature. For each feature point category, statistical metrics such as mean and variance are used to calculate the statistical characteristics of that category, forming a structured feature description. The structured feature descriptions of each category are combined to construct a structured environmental feature dataset, with each data sample containing feature descriptions from multiple categories. The CART decision tree algorithm is used to train the environmental feature dataset to establish an environmental perception model for determining the type and state of the current environment.

[0030] For example, a city patrol robot collects distance data using a range sensor. This raw data is affected by the robot's position and device errors, resulting in noise. The Kalman filter effectively removes this noise through two stages: prediction and update. For example, if the robot measures a distance of 10.2 meters to an obstacle at a given moment, the Kalman filter, by incorporating the robot's previous position and measurement uncertainty, may output a more accurate distance estimate of 10.18 meters. The SIFT algorithm is particularly important for extracting stable feature points in complex environments. For example, a city rescue robot needs to identify obstacles under varying lighting conditions and viewing angles. The SIFT algorithm can extract feature descriptors from the obstacle's texture that are invariant to scale, rotation, and brightness. For example, the bark of an oak tree may produce hundreds of feature points, each of which is described by a 128-dimensional vector describing its local gradient information. These feature points are reliably identified across seasons and time of day, providing a reliable reference for robot navigation. The K-means clustering algorithm helps robots understand the structure of their environment. For example, a city patrol robot needs to distinguish between buildings, vehicles, and pedestrians. By clustering SIFT feature points, the following categories may be formed: tall vertical structures (buildings), horizontally extending structures (roads), small moving objects (pedestrians), etc. Assuming the number of clusters K = 5, after algorithm iterations, these five categories of environmental features may be obtained, each containing hundreds of similar feature points. Statistical feature descriptions provide quantitative metrics for each environmental category. For feature points in the "building" category, their average height, density, and texture complexity can be calculated. For example, if the average height is 20 meters, the density is 0.5 feature points per square meter, and the texture complexity (based on gradient variance) is 0.8. These values constitute a structured feature description, enabling the robot to quickly determine the nature of its surroundings. The CART decision tree algorithm builds an environmental perception model based on these structured features. For example, the model may learn the following rule: if the average height of feature points is greater than 15 meters and the density is greater than 0.3 / ㎡, the area is considered to be a high-rise building area; if the average height is less than 2 meters and the movement speed is greater than 1m / s, the area is considered to be densely populated with pedestrians. This enables robots to analyze their surroundings in real time and make appropriate behavioral decisions, such as slowing down to avoid collisions in high-rise areas and increasing vigilance in pedestrian areas. Through the synergistic effect of this series of algorithms, biomimetic robots are able to extract a high-level understanding of the environment from raw sensor data. This not only improves the robot's adaptability in complex and dynamic environments but also lays the foundation for subsequent path planning and task execution. For example, urban patrol robots can quickly respond to unusual situations, improving urban safety. This enhanced environmental perception capability enables biomimetic robots to play a greater role in various application scenarios, driving the advancement of artificial intelligence and robotics.

[0031] Step S103: Based on the structured environmental feature data set, a Gaussian mixture model is used to construct an environmental model. If the environmental change exceeds a preset threshold, the parameters of the environmental model are updated to reflect the latest environmental status.

[0032] Based on the biomimetic robot's structured environmental feature dataset, principal component analysis is used to reduce the dimensionality of the environmental features, generating a reduced environmental feature vector. The K-means clustering algorithm is then used to cluster the environmental features based on the reduced dimensionality vector, generating a clustering result. Based on the clustering result, a Gaussian mixture model is used to construct a model for each cluster, generating an initial environmental model. Based on the initial environmental model, the model parameters are optimized using recursive least squares to obtain the optimized environmental model parameters. Current environmental data is acquired in real time, and an environmental feature vector is extracted from it. The Euclidean distance is calculated between the vectors in the historical environmental feature dataset, and the minimum distance is obtained as the degree of environmental change. If the degree of environmental change exceeds a preset threshold, the environmental model update mechanism is triggered, and the model parameters are dynamically adjusted using recursive least squares based on the latest environmental feature data. The current environmental feature vector is matched with the adjusted environmental model parameters, generating the cosine similarity as the environmental similarity. If the similarity exceeds a preset threshold, the model is considered to have adapted to the latest environmental conditions.

[0033] For example, for a city patrol robot, the raw environmental features may contain hundreds of dimensions, such as building height, road width, and pedestrian density. Principal component analysis (PCA) can condense these features down to 10 principal components, each of which is a linear combination of the original features. For example, the first principal component might reflect the overall size of the city, while the second might represent commercial activity. This dimensionality reduction not only reduces data storage and computational overhead but also highlights the essential characteristics of the environment. The K-means clustering algorithm classifies the environment in this reduced feature space. Assuming the city environment is divided into five categories, the algorithm might produce the following clustering results: high-density commercial areas, low-density residential areas, industrial parks, parks and green spaces, and transportation hubs. Each category has its own characteristic center. For example, the characteristic centers of high-density commercial areas might be high building density, high pedestrian flow, and complex road networks. This clustering provides the robot with a macroscopic understanding of the city's structure. A Gaussian mixture model finely models each cluster, capturing the probability distribution of environmental characteristics. For example, in a high-density commercial area, the model might contain multiple Gaussian components corresponding to different commercial subtypes, such as shopping malls, office complexes, and restaurant districts. Each Gaussian component has its own mean and covariance matrix, describing the characteristic distribution of that subtype. This modeling approach can reflect the diversity and complexity of the environment. Recursive least squares is used to dynamically optimize environmental model parameters, enabling the model to adapt to gradual changes in the environment. For example, as a city develops, an area may gradually transform from a low-density residential area to a medium-density mixed area. Recursive least squares continuously adjusts the parameters of the Gaussian mixture model, such as the mean vector and covariance matrix, based on new observations, ensuring that the model always reflects the current environmental state. The degree of environmental variability is calculated by comparing the Euclidean distance between the current feature vector and historical data. Suppose a robot, during a patrol, discovers a sudden increase in pedestrian density and a significant change in building height in an area. This may cause the Euclidean distance from historical data to exceed a preset threshold, such as 0.5. At this point, the environmental model update mechanism is triggered, and the system reassesses the area's classification, potentially reclassifying it from a low-density residential area to an emerging commercial area under development. Environmental similarity is calculated using cosine similarity, measuring the degree of agreement between the current environmental feature vector and the model's prediction. For example, if the cosine similarity between the observed features of an area and a model of a high-density commercial district reaches 0.9 or higher, the robot can be confident that it is currently in a typical commercial area. This real-time environmental recognition capability enables the robot to quickly adjust its behavioral strategies, such as reducing speed and increasing vigilance in crowded areas. Through this series of environmental modeling and dynamic update mechanisms, the biomimetic robot is able to build a stable and flexible environmental perception system. This not only improves the robot's adaptability in complex urban environments but also provides a solid foundation for intelligent decision-making and mission planning.For example, patrol robots can adjust patrol frequency and focus based on the characteristics of different areas, focusing on preventing theft in commercial areas and paying attention to safety hazards in residential areas. This intelligent environmental perception and response mechanism greatly improves the efficiency and precision of urban management.

[0034] Step S104: Based on the environmental model, a time series prediction algorithm is used to predict future environmental change trends. If the deviation between the prediction result and the current environmental state exceeds a preset range, a model retraining mechanism is triggered.

[0035] Historical environmental data from the environmental model is obtained to construct a time series dataset, which includes environmental state parameters (such as obstacle locations, target states, and terrain features) and corresponding timestamps. Preprocessing of the time series data includes missing value filling, outlier handling, and data normalization. Training samples are constructed using a sliding window method. Each sample contains a sequence of environmental states from the past as input and environmental states from the future as output. The long short-term memory (LSTM) neural network algorithm is used to train these training samples and establish an environmental state prediction model. The model's prediction performance is optimized by adjusting hyperparameters such as the number of LSTM layers, number of hidden units, and learning rate. Current environmental state data is obtained online, using environmental sensors to collect real-time values for parameters such as temperature, humidity, and air pressure. The current environmental state data is input into the prediction model to generate predicted environmental states for the future. The deviation between the current actual environmental state and the predicted value is calculated using common error metrics such as mean squared error (MSE) or mean absolute error (MAE). The deviation is compared with a preset threshold, which can be set based on the fluctuation of historical data, such as when the deviation exceeds twice the historical data fluctuation range. If the deviation exceeds the threshold, the model retraining mechanism is triggered. The latest environmental status data is added to the training set, and the environmental status prediction model is incrementally trained using online learning to adapt the model to the latest environmental change trends.

[0036] For example, building an environmental prediction model for a biomimetic robot begins with the collection and organization of historical environmental data. This data contains a wealth of environmental state information, such as obstacle locations, target states, and terrain features. For example, in an urban environment, obstacles might be moving vehicles or pedestrians, target states might be changing traffic lights, and terrain features might include road slope and pavement material. This data is arranged chronologically, forming a multidimensional time series dataset. Data preprocessing is a critical step in ensuring the quality of model training. In real-world applications, sensor failures or signal interference may lead to missing data. For example, in severe weather conditions, some optical sensors may not function properly. In such cases, interpolation methods, such as linear interpolation or more complex spline interpolation, can be used to fill missing values. Handling outliers is equally important. For example, sudden extreme temperature readings may be caused by sensor failure and require statistical methods to identify, remove, or correct them. Data normalization helps balance the scale differences of different features, allowing data with different dimensions, such as temperature (which may range from -10 to 40°C) and humidity (0-100%), to be compared and calculated on the same scale. The sliding window method is an effective way to construct training samples. Assuming a 24-hour input window and a 6-hour prediction window, each training sample contains the environmental state sequence for the past 24 hours as input and the environmental state for the next 6 hours as output. This approach allows the model to learn short-term and medium-term patterns of environmental change. The Long Short-Term Memory (LSTM) network was chosen based on its ability to capture long-term dependencies. This property is particularly important in environmental forecasting, as some environmental changes can be influenced by events that occurred hours or even days prior. For example, changes in urban traffic flow may be related to large-scale events that occurred several days prior. By adjusting the LSTM structure, such as adding layers to capture more complex patterns or increasing the number of hidden units to increase model capacity, forecasting performance can be optimized. Adjusting the learning rate influences the convergence speed and stability of the model. Real-time environmental state data provides the model with the latest input. For example, a temperature sensor may record readings every minute, while a humidity sensor may record every 5 minutes. After preprocessing, this data is input into the forecasting model to generate a forecast of environmental conditions for the next 6 hours. Calculating and evaluating the forecast deviation triggers the dynamic updating of the model. Assume the predicted temperature is 25°C and the actual observed value is 28°C, then the mean square error (MSE) is 9. If the temperature fluctuation range of historical data is typically within ±2°C, the threshold can be set to 4°C. When the prediction deviation exceeds this threshold, it indicates that the current model may no longer accurately reflect environmental trends and needs to be updated. The model's online learning mechanism ensures its continuous adaptability to environmental changes. For example, if a city embarks on a large-scale greening project, the temperature and humidity characteristics of the environment may gradually change.By continuously incorporating new observations into the training set and incrementally training the model, the prediction results can maintain a high level of accuracy. This dynamic update mechanism enables biomimetic robots to maintain efficient decision-making capabilities in ever-changing environments, laying the foundation for their application in complex scenarios.

[0037] Step S105: Based on the prediction results, combined with the current state and target task of the bionic robot, a reinforcement learning algorithm is used to generate an optimal decision strategy.

[0038] A1. Obtain the bionic robot's current state information and target task requirements, extract features, and vectorize them as state inputs for the reinforcement learning algorithm. A2. Based on the state inputs and a predefined action space, the Q-learning algorithm is used to evaluate and select actions. In the Q-learning algorithm, the state space is the set of all possible states of the bionic robot, and the action space is the set of all executable actions of the robot. The reward function is designed based on the degree of proximity between the state after each action and the target state. A3. By continuously trying actions and observing state changes and reward feedback, the Bellman equation is used to update the Q-value table, allowing the Q-value to converge and obtain the optimal action-value function. A4. At each decision moment, the Q-value table is consulted based on the current state, and the action with the largest Q-value is selected as the optimal decision, forming an optimal decision sequence. A5. The optimal decision sequence is converted into control instructions, which drive the bionic robot's actuators to perform the corresponding actions, and the post-execution state feedback is obtained through sensors. A6. Determine whether the current state meets the target task requirements. If so, output the optimal decision strategy and end. If not, use the current state as the new initial state and return to step 2 to iterate the decision again until the goal is achieved.

[0039] For example, the core of a biomimetic robot's decision-making system lies in reinforcement learning algorithms, with QLearning being a widely used approach. First, the robot's current state information and target task requirements are obtained. For example, for an indoor service robot, its state may include location coordinates, battery level, and current task progress. This information needs to be converted into a vector representation through feature extraction for algorithm processing. Assuming the robot's position is represented by (x, y) coordinates, the battery level as a percentage, and the task progress as completion, a possible state vector might be [2.5, 3.7, 80%, 60%]. The definition of the action space directly impacts the robot's behavioral capabilities. For an indoor service robot, actions may include moving forward, backward, turning, and grasping objects. Each action needs to be quantified. For example, moving forward can be defined as [1, 0], backward as [-1, 0], turning left as [0, -1], and turning right as [0, 1]. This representation enables the robot to move precisely on a two-dimensional plane. The design of the reward function is key to the QLearning algorithm. It determines how the robot evaluates its own actions. For a service robot, completing a task earns positive rewards, while colliding with obstacles or running out of power earns negative rewards. For example, successfully delivering an item to its destination might earn a reward of +100, while each movement consumes a reward of -1, encouraging the robot to find the shortest path. Updating the Q-value table is central to the learning process. Initially, the Q-value table may be randomly initialized. As the robot attempts different actions and observes the results, the Q-values gradually converge. For example, if the robot discovers that turning right brings it closer to the goal faster, the Q-value for turning right in a particular state will gradually increase. The optimal decision sequence is generated based on the converged Q-value table. Suppose the robot needs to move from room A to room B. Querying the Q-value table might yield a decision sequence like: [Forward, Forward, Turn Right, Forward, Turn Left, Forward]. This sequence is converted into specific control instructions, such as motor speed and direction. After executing the action, the robot obtains new state information through various sensors. For example, a lidar can update the map of the surrounding environment, an odometry can accurately calculate the distance traveled, and a camera can determine whether the target location has been reached. This feedback is used to evaluate the effectiveness of the current decision and serves as input for the next round of decision-making. The entire process is iterative, with the robot continually adjusting its decisions to adapt to environmental changes and task requirements. This adaptive capability enables biomimetic robots to efficiently complete tasks in complex and dynamic environments, demonstrating intelligent behavior similar to that of living creatures.

[0040] Step S106: Generate a control instruction through the decision strategy. If the control instruction conflicts with the current execution state, call the robustness control module to adjust the instruction priority. The robustness control module adapts to environmental changes by dynamically adjusting control parameters to ensure the stability and adaptability of instruction execution.

[0041] Initial control instructions are generated based on the decision-making strategy. Sensors are used to acquire current system execution status information, such as environmental parameters and equipment operating parameters. The collected status information undergoes preprocessing, including filtering and normalization, to improve data quality. The processed status information is compared and analyzed with the initial control instructions to determine whether there are logical or physical conflicts, such as instructions exceeding the equipment's operating range or timing inconsistencies between instructions. If a conflict exists, the robust control module is triggered to handle the conflict. The robust control module acquires relevant information, including control instructions, execution status, and environmental changes. With system stability as the control objective, it dynamically adjusts the controller output using a PID control algorithm, ensuring that the system state variables quickly track the setpoints under disturbances. The PID controller's input is the state variable deviation, and its output is the actuator drive signal. Building on PID control, adaptive control algorithms, such as the MIT rule, are introduced to online estimate the controlled object parameters and adjust the PID parameters in real time, improving system adaptability. The adaptive rate can be dynamically adjusted based on the system state change trend. Control instructions are ranked based on the optimized control parameters output by the robust control module. Instruction priority evaluation metrics include task urgency, execution timeliness, and impact on system stability. A weighted summation method is used to calculate the overall priority. Instruction timestamp information is used to determine the execution order for instructions with the same priority. The execution module receives the optimized control instruction sequence and determines the execution order based on instruction priority and timestamp. For critical instructions with high priority and strong timeliness, an interrupt mechanism is used to preempt execution. The resource scheduling strategy utilizes a priority preemption algorithm to prioritize the resources required by high-priority instructions.

[0042] For example, initial control instructions are first generated based on the decision strategy. These instructions may include the robot's movement direction, speed, and force. For example, for a transport robot, the initial instruction might be "move forward 5 meters at a speed of 0.5 meters per second." Next, various sensors are used to acquire environmental and equipment operating parameters. These sensors may include cameras, lidar, and accelerometers. For example, in the case of a transport robot, its lidar might detect an obstacle 2 meters ahead, and its force sensor might sense a 50 kg load. The acquired raw data requires preprocessing to improve its quality. For example, accelerometer data can be low-pass filtered to remove high-frequency noise, and camera images can be normalized to accommodate varying lighting conditions. This processed data is more reliable and facilitates subsequent decision-making and control. The system then compares and analyzes this processed state information with the initial control instructions. In the case of the transport robot, the system might detect a conflict between the "move forward 5 meters" instruction and the "obstacle 2 meters ahead" state information. This conflict could lead to a collision, triggering the robust control module for resolution. The core of the robust control module is the PID control algorithm. For a transport robot, a PID controller might use the distance to an obstacle as input and output a speed adjustment signal for the robot. For example, when the robot is far from an obstacle, the PID controller might output a high positive velocity. As it approaches the obstacle, the output decreases or even becomes negative, causing the robot to decelerate or stop. To improve system adaptability, adaptive control algorithms such as the MIT rule are introduced. This allows PID parameters to automatically adjust to environmental changes. For example, when the robot enters rough terrain from flat ground, the MIT rule might increase the weight of the PID controller's integral term to better overcome the disturbances caused by terrain changes. The optimized control instructions need to be prioritized. Based on a pre-defined evaluation metric system, multi-dimensional evaluation data is obtained for each initial control instruction, including task urgency, execution timeliness, and impact on system stability. Principal component analysis is used to reduce the dimensionality of this multi-dimensional evaluation data, extracting a comprehensive evaluation score that serves as the optimized control parameter for each instruction. Weight coefficients for each evaluation metric are set according to business rules, and a weighted summation method is used to calculate the comprehensive priority score for each instruction. The initial control instructions are sorted in descending order based on the comprehensive priority score to obtain the optimized instruction sequence. The command sequence is checked to see if there are any instructions with the same priority score. If so, they are sorted again based on the chronological order of their timestamps to determine the priority of the instructions. For a transport robot, the instruction "avoid collision" might be given the highest priority, followed by "maintain balance," and then "complete the task on time." This ordering ensures that the robot prioritizes safety while completing the task. Finally, the execution module executes the instructions based on their priority and timestamps. High-priority instructions may interrupt the current task.For example, if a transport robot suddenly detects a person approaching during movement, an "avoid collision" command is immediately executed, pausing the current transport task. This mechanism ensures the robot can flexibly respond to various emergencies, improving overall safety and efficiency. Through this series of steps, the bionic robot can operate stably and reliably in complex and changing environments, demonstrating adaptability and intelligence similar to living organisms. This control system design not only improves the robot's performance but also enhances its practicality and reliability in various application scenarios.

[0043] The present invention provides a bionic robot control system based on AI learning, which mainly includes:

[0044] Environmental data acquisition module, used to obtain dynamic environmental data including obstacle location, target status and terrain characteristics, and collect and transmit real-time data to the data processing module through the sensor network;

[0045] The data processing module is used to remove noise and redundant information from dynamic environmental data, extract key features, and generate a structured environmental feature data set;

[0046] The environmental modeling module is used to build an environmental model based on a structured environmental feature dataset using a Gaussian mixture model. If the environmental change exceeds a preset threshold, the parameters of the environmental model are updated to reflect the latest environmental status.

[0047] The prediction analysis module is used to predict future environmental change trends based on the environmental model using a time series prediction algorithm. If the deviation between the prediction result and the current environmental state exceeds a preset range, the model retraining mechanism is triggered;

[0048] The decision generation module is used to generate the optimal decision strategy based on the prediction results, combined with the current state of the bionic robot and the target task, using the reinforcement learning algorithm;

[0049] The robustness control module is used to generate control instructions through decision-making strategies. If the control instructions conflict with the current execution status, the robustness control module is called to adjust the instruction priority. The robustness control module adapts to environmental changes by dynamically adjusting control parameters to ensure the stability and adaptability of instruction execution.

[0050] The above content is merely an example and explanation of the structure of the present invention. Those skilled in the art may make various modifications or additions to the described specific embodiments or replace them in a similar manner. As long as they do not deviate from the structure of the invention or exceed the scope defined by the claims, they should all fall within the scope of protection of the present invention.

Claims

1. A bionic robot control method based on AI learning, characterized in that: The method comprises: The sensors acquire dynamic environmental data including obstacle positions, target status and terrain features in real time, and transmit the real-time data to the data processing module through the sensor network; For dynamic environmental data, remove noise and redundant information, extract key features, and generate a structured environmental feature dataset; Based on a structured environmental feature dataset, a Gaussian mixture model is used to build an environmental model. If the environmental change exceeds a preset threshold, the parameters of the environmental model are updated to reflect the latest environmental status. Based on the environmental model, a time series prediction algorithm is used to predict future environmental change trends. If the deviation between the predicted result and the current environmental state exceeds a preset range, the model retraining mechanism is triggered; Based on the prediction results, the reinforcement learning algorithm is used to generate the optimal decision strategy in combination with the current state and target task of the bionic robot; Generate control instructions through decision-making strategies. If the control instructions conflict with the current execution status, the robustness control module is called to adjust the instruction priority. The robustness control module adapts to environmental changes by dynamically adjusting control parameters to ensure the stability and adaptability of instruction execution. The environmental model is constructed using a Gaussian mixture model based on a structured environmental feature dataset. If the environmental change exceeds a preset threshold, the parameters of the environmental model are updated to reflect the latest environmental status, including: Obtaining structured environmental feature datasets; The principal component analysis method is used to reduce the dimension of the environmental feature data set to obtain the reduced dimension environmental feature vector; According to the environmental feature vector after dimensionality reduction, the K-means clustering algorithm is used to cluster and obtain the environmental feature clustering results; For each environmental feature clustering result, a Gaussian mixture model is used to build a model to obtain the initial environmental model; The recursive least square method is used to optimize the parameters of the initial environmental model to obtain the optimized environmental model parameters; Obtain current environment data and extract current environment feature vector; The Euclidean distance between the current environmental feature vector and the historical environmental feature vector in the environmental feature dataset is calculated, and the minimum distance value is determined as the degree of environmental change; If the degree of environmental change exceeds a preset threshold, the environmental model update mechanism is triggered, and the environmental model parameters are dynamically adjusted using the recursive least squares method based on the latest environmental feature data; The adjusted environment model parameters are matched with the current environment feature vector to obtain the cosine similarity; If the cosine similarity is higher than the preset threshold, it is determined that the environmental model has adapted to the latest environmental state.

2. The method according to claim 1, characterized in that The method of acquiring dynamic environmental data including obstacle positions, target states, and terrain features in real time through sensors and transmitting the real-time data to a data processing module through a sensor network includes: Obtain the lidar point cloud data and camera image data collected by the bionic robot; Perform filtering, denoising and downsampling on the LiDAR point cloud data to obtain pre-processed LiDAR point cloud data; Performing filtering, denoising and histogram equalization on the camera image data to obtain preprocessed camera image data; The pre-processed LiDAR point cloud data is input into the SLAM module, and the ICP algorithm is used to locate the robot position in real time and build an environmental map; The pre-processed camera image data is input into the target detection module, and the YOLO algorithm is used to detect and track the target object in real time; Based on the environment map constructed by the SLAM module, the lidar point cloud data is rasterized to generate a two-dimensional occupancy grid map; Project the target object bounding box information output by the target detection module onto the two-dimensional occupancy grid map and update the target object position information; The SVM algorithm is used to classify the grids in the two-dimensional occupancy grid map, obtain the grid passability probability, and extract the grid's terrain features; Convert the two-dimensional occupancy grid map into a cost map based on the grid's passability probability and terrain characteristics; The A* search algorithm is used to search for the optimal navigation path, with the robot's current position as the starting point, the target position as the end point, the two-dimensional occupancy grid map as the environment model, and the cost map as the heuristic function; Guide the robot to navigate autonomously according to the optimal navigation path, and update the environmental map and obstacle avoidance information in real time.

3. The method according to claim 1, characterized in that The method for removing noise and redundant information from dynamic environmental data, extracting key features, and generating a structured environmental feature dataset includes: Acquire dynamic environmental data, and use Kalman filter algorithm to filter the dynamic environmental data to obtain filtered environmental data; For the filtered environmental data, the SIFT algorithm is used to extract key feature points to obtain feature descriptors that are invariant to scale, rotation, and brightness. According to the extracted SIFT feature points, the K-means clustering algorithm is used to cluster the SIFT feature points to obtain feature point sets of different categories, where each category represents an environmental feature; For each feature point category, the statistical characteristics of the category are calculated using mean and variance statistical indicators to form a structured feature description; The structured feature descriptions of each category are combined to construct a structured environmental feature dataset, where each data sample contains feature descriptions of multiple categories; The CART decision tree algorithm is used to train the structured environmental feature dataset to obtain the environmental perception model; During the operation of the bionic robot, the current environmental data is obtained, and the environmental perception model is used to analyze the current environmental data to determine the type and status of the current environment.

4. The method according to claim 1, characterized in that The time series prediction algorithm is used to predict future environmental change trends based on the environmental model. If the deviation between the prediction result and the current environmental state exceeds a preset range, the model retraining mechanism is triggered, including: Obtain historical environmental status data from the environmental model. The historical environmental status data includes obstacle locations, target states, terrain feature parameters, and corresponding timestamps, and construct a time series dataset. Preprocessing of time series datasets, including missing value filling, outlier handling, and data normalization; The sliding window method is used to construct training samples. Each training sample contains a sequence of environmental states over a period of time as input and an environmental state over a period of time in the future as output. The long short-term memory neural network algorithm is used to train the training samples and establish an environmental state prediction model; Obtain the actual environmental status data at the current moment, input the actual environmental status data into the environmental status prediction model, and obtain the predicted environmental status data for a period of time in the future; Calculate the deviation between the actual environmental state data and the predicted environmental state data. If the deviation exceeds the preset threshold, the model retraining mechanism is triggered; The latest environmental status data is added to the training samples, and the environmental status prediction model is incrementally trained using online learning to make it adapt to the latest environmental change trends.

5. The method according to claim 1, characterized in that Based on the prediction results, combined with the current state and target task of the bionic robot, the reinforcement learning algorithm is used to generate the optimal decision strategy, including: A1. Obtain the current state information and target task requirements of the bionic robot, extract features from the state information and target task requirements, and vectorize them to obtain a state feature vector. A2. Use the state feature vector as the state input of the reinforcement learning algorithm. Combined with the preset action space, the Q-learning algorithm is used to evaluate the actions and obtain the Q value of each action. In the Q-learning algorithm, the state space is the set of all states of the bionic robot, the action space is the set of all executable actions of the bionic robot, and the reward function is designed based on the degree to which the state after each action is executed is relative to the target state; A3. By continuously trying actions and observing state changes and reward feedback, the Bellman equation is used to update the Q-value table so that the Q-value converges to obtain the optimal action-value function. A4. At each decision moment, query the Q-value table based on the current state and select the action with the largest Q-value as the optimal decision, forming the optimal decision sequence; A5. Convert the optimal decision sequence into control instructions, drive the actuators of the bionic robot to perform the corresponding actions, and obtain status feedback after execution through sensors; A6. Determine whether the state feedback meets the target task requirements. If so, output the optimal decision strategy and end. If not, the current state is used as the new initial state and the process returns to step A2 to iterate the decision again until the goal is achieved.

6. The method according to claim 1, characterized in that The control instructions are generated through the decision-making strategy. If the control instructions conflict with the current execution state, the robustness control module is called to adjust the instruction priority. The robustness control module adapts to environmental changes by dynamically adjusting control parameters to ensure the stability and adaptability of instruction execution, including: Obtaining initial control instructions and current system execution status information, including environmental parameters and equipment operating parameters; Preprocessing the current system execution status information to obtain processed status information; Determine whether the processed state information conflicts with the initial control instruction. If so, trigger the robustness control module. The robust control module obtains the initial control instructions and processed state information and environmental change information, and outputs the optimized control parameters through the PID control algorithm. The input of the PID controller is the state variable deviation, and the output is the actuator drive signal; Based on the PID control algorithm, an adaptive control algorithm is used to estimate the parameters of the controlled object online, correct the PID parameters in real time, and dynamically adjust the adaptive rate according to the trend of system state changes; The initial control instructions are sorted according to the optimized control parameters to obtain a control instruction sequence. The instruction priority evaluation indicators include task urgency, execution timeliness and impact on system stability. The weighted summation method is used to calculate the comprehensive priority. The execution module receives a sequence of control instructions and determines the execution order based on the instruction priority and timestamp. For critical instructions with high priority and strong timeliness, the interrupt mechanism is used to preempt execution. The resource scheduling strategy adopts a priority preemption algorithm to give priority to meeting the resources required by high-priority instructions.

7. A bionic robot control system based on AI learning, characterized in that: The system comprises: Environmental data acquisition module, used to obtain dynamic environmental data including obstacle location, target status and terrain characteristics in real time through sensors, and transmit real-time data to the data processing module through the sensor network; The data processing module is used to remove noise and redundant information from dynamic environmental data, extract key features, and generate a structured environmental feature data set; The environmental modeling module is used to build an environmental model based on a structured environmental feature dataset using a Gaussian mixture model. If the environmental change exceeds a preset threshold, the parameters of the environmental model are updated to reflect the latest environmental status. The prediction analysis module is used to predict future environmental change trends based on the environmental model using a time series prediction algorithm. If the deviation between the prediction result and the current environmental state exceeds a preset range, the model retraining mechanism is triggered; The decision generation module is used to generate the optimal decision strategy based on the prediction results, combined with the current state of the bionic robot and the target task, using the reinforcement learning algorithm; The robustness control module is used to generate control instructions through decision-making strategies. If the control instructions conflict with the current execution status, the robustness control module is called to adjust the instruction priority. The robustness control module adapts to environmental changes by dynamically adjusting control parameters to ensure the stability and adaptability of instruction execution. The environmental model is constructed using a Gaussian mixture model based on a structured environmental feature dataset. If the environmental change exceeds a preset threshold, the parameters of the environmental model are updated to reflect the latest environmental status, including: Obtaining structured environmental feature datasets; The principal component analysis method is used to reduce the dimension of the environmental feature data set to obtain the reduced dimension environmental feature vector; According to the environmental feature vector after dimensionality reduction, the K-means clustering algorithm is used to cluster and obtain the environmental feature clustering results; For each environmental feature clustering result, a Gaussian mixture model is used to build a model to obtain the initial environmental model; The recursive least square method is used to optimize the parameters of the initial environmental model to obtain the optimized environmental model parameters; Obtain current environment data and extract current environment feature vector; The Euclidean distance between the current environmental feature vector and the historical environmental feature vector in the environmental feature dataset is calculated, and the minimum distance value is determined as the degree of environmental change; If the degree of environmental change exceeds a preset threshold, the environmental model update mechanism is triggered, and the environmental model parameters are dynamically adjusted using the recursive least squares method based on the latest environmental feature data; The adjusted environment model parameters are matched with the current environment feature vector to obtain the cosine similarity; If the cosine similarity is higher than the preset threshold, it is determined that the environmental model has adapted to the latest environmental state.

Images