A logistics robot control method and system based on online transfer learning

Abstract

This application relates to the field of logistics robot technology, and in particular to a logistics robot control method and system based on online transfer learning. It includes: collecting environmental state data through the logistics robot's sensor array, generating a state vector based on the environmental state data; inputting the state vector in parallel to two processing channels of a preset control algorithm, and outputting motion control commands for the logistics robot; updating the state vector based on the execution results of the motion control commands, generating learning signal parameters, and setting an optimization learning strategy for the control algorithm based on the learning signal parameters; and achieving functional decoupling of high-level decision-making and rapid adaptation by constructing a parallel dual-pathway architecture of a brain-decision network and a cerebellum-adaptive network. The brain-decision network outputs basic motion commands based on cross-environment task planning and strategy transfer; the cerebellum-adaptive network focuses on handling dynamic disturbances in the logistics scenario, providing real-time fine-tuning.

Description

Technical Field

[0001] This application relates to the field of logistics robot technology, and in particular to a logistics robot control method and system based on online transfer learning. Background Technology

[0002] Existing logistics robot control algorithms exhibit the following significant shortcomings in real-world dynamic environments: the parameters of the trained policy network are fixed, making it difficult to cope with frequent dynamic changes in logistics scenarios. For example, when there are sudden changes in cargo weight (empty → fully loaded), changes in ground material (epoxy flooring → metal grating), adjustments to shelf layout, or the appearance of temporary obstacles, the system needs to collect a large amount of data for offline retraining, resulting in long business interruptions and low operational efficiency.

[0003] Traditional deep reinforcement learning algorithms lack the ability to transfer across warehouses and vehicle models. In new logistics environments (such as migrating from e-commerce warehouses to cold chain warehouses), algorithms need to be trained from scratch, requiring over 100,000 interactions and deployment cycles lasting several weeks, making it difficult to meet the logistics industry's needs for rapid deployment and flexible scheduling. Significant multi-timescale challenges exist in logistics scenarios: the frequency difference between visual navigation signals (10-30Hz) and motor control signals (100-1000Hz) is enormous. Direct end-to-end training leads to policy instability, resulting in issues such as motion jitter and trajectory oscillations, increasing mechanical wear and energy consumption.

[0004] Existing online learning mechanisms rely on reward feedback, resulting in a significant delay (typically >1 second) between environmental changes and policy updates. In highly dynamic scenarios such as forklift obstacle avoidance and narrow passage, this inability to compensate for disturbances in real time leads to high collision risks and low task success rates. Summary of the Invention

[0005] The purpose of this application is to provide a logistics robot control method and system based on online transfer learning to solve the above-mentioned technical problems, aiming to achieve autonomous adaptation and continuous optimization of logistics robots in dynamic environments.

[0006] In some embodiments of this application, a dual-pathway architecture is constructed that parallels a brain-based decision-making network (high-level strategy) and a cerebellum-based adaptive network (low-level compensation) to achieve functional decoupling between high-level decision-making and rapid adaptation. The brain-based decision-making network outputs basic motor commands based on cross-environment task planning and strategy transfer; the cerebellum-based adaptive network focuses on handling dynamic disturbances in logistics scenarios (such as changes in cargo weight and ground friction coefficient), providing real-time fine-tuning. This ensures the macro-optimal nature of the strategy and the system's rapid response capability to sudden changes.

[0007] In some embodiments of this application, by monitoring changes in the distribution of input features in real time, the update strategy of each layer of the brain's decision network is dynamically determined. Through lightweight fine-tuning, new scenarios can be quickly adapted, improving the efficiency of adaptation to different environments, while preventing catastrophic forgetting and maintaining the memory of the learned environment.

[0008] In some embodiments of this application, by decoupling high-level decision transfer and low-level dynamic compensation, the brain network focuses on semantic understanding and policy transfer across environments, without needing to fit complex and ever-changing environmental dynamics; the cerebellum module specifically handles high-frequency perturbation compensation, with a clear learning objective, improving convergence efficiency and stability. By monitoring changes in environmental distribution and dynamically adjusting network parameters, the system can adapt to new warehouse environments within minutes, without requiring extensive offline data collection and retraining, thus improving deployment efficiency.

[0009] In some embodiments of this application, a logistics robot control method based on online transfer learning is provided, including: The environmental state data is collected by the sensing array of the logistics robot, and a state vector is generated based on the environmental state data. The state vector input is fed in parallel to two processing channels of a preset control algorithm, and the motion control commands for the logistics robot are output. The state vector is updated based on the execution result of the action control command, and learning signal parameters are generated. The optimization learning strategy of the control algorithm is set based on the learning signal parameters. The control algorithm includes: Brain decision-making network and cerebellum adaptive network; The two processing channels include: the brain decision-making channel and the cerebellum compensation channel.

[0010] In some embodiments of this application, the establishment of the preset control algorithm includes: Obtain historical environmental data; Multiple environmental characteristic indicators are selected based on the historical environmental data; A scene semantic model and multiple environmental scenarios are established based on all the aforementioned environmental feature indicators; Establish the expected action commands for each environmental scenario in sequence; The expected action command includes: expected linear velocity and expected angular velocity; The brain decision-making network is constructed based on all expected action instructions and the scene semantic model; Establish the brain connection channel that connects to the brain decision-making network; Select multiple main equipment indicators; A threshold activation function and a weight table are established based on all the aforementioned device indicators; The cerebellum adaptive network is constructed based on the threshold activation function and the weight table. Establish the cerebellar compensation channel that is connected to the cerebellar adaptive network.

[0011] In some embodiments of this application, the output motion control commands for the logistics robot include: Visual and point cloud data are obtained based on the state vector; The visual and point cloud data are input into the brain connection channel, and the scene semantic model extracts global scene semantics from the visual and point cloud data. Output basic action instructions based on the global scene semantics; The ontology perception signal is obtained based on the state vector; The proprioceptive signal is input into the cerebellar compensation channel and converted into a sparse pulse sequence through the threshold activation function. The sparse pulse sequence is mapped to the weight table through a hash mapping input space, and an instantaneous compensation action is generated. The motion control command is generated based on the basic motion command and the instantaneous compensation motion.

[0012] In some embodiments of this application, generating the motion control command includes: An environmental fluctuation quantity is generated based on the environmental state data. The weighting coefficient of the instantaneous compensation action is set according to the amount of environmental fluctuation; Based on the weighting coefficients, the basic action command and the instantaneous compensation action are vector-synthesized; The motion control command is generated based on the synthesis result.

[0013] In some embodiments of this application, the generated learning signal parameters include: Establish a timeline that includes multiple consecutive moments; Generate the expected environmental state for the next moment at the current moment; When the logistics robot finishes executing the action control command at the current moment, the sensing array acquires the actual observation state at the next moment; The sensory prediction error is calculated based on the expected environmental state and the actual observed state. Generate prediction error signal; Obtain the current task completion rate and calculate the sparse reward based on the task completion rate; Generate task reward signals; Learning signal parameters are generated based on the prediction error signal and the task reward signal.

[0014] In some embodiments of this application, the optimization learning strategy of the setting control algorithm includes: Constructing a reward feedback channel for a cerebellar adaptive network; Obtain the prediction error signal from the learned signal parameters; The prediction error signal is input into the reward feedback channel to generate the weight update parameters of the cerebellum adaptive network; Obtain the task reward signal from the learning signal parameters; Based on the task reward signal, determine whether to generate a migration update instruction for the brain decision network; The optimization learning algorithm is set based on the weight update parameters and migration update instructions.

[0015] In some embodiments of this application, the step of determining whether to generate a migration update instruction for the brain decision network includes: Generate a cumulative reward value based on the task reward signal; If the accumulated reward value exceeds the preset reward value threshold, a migration update instruction will be generated. The migration update instructions include: Based on the collected environmental interaction data, the distribution changes of environmental features are generated; The migration and update range of the brain decision network is selected based on the distribution of changes in the environmental characteristics. Perform lightweight adjustments based on the migration update scope.

[0016] In some embodiments of this application, a logistics robot control system based on online transfer learning is provided, comprising: The data processing unit is used to collect environmental state data through the sensing array of the logistics robot and generate a state vector based on the environmental state data. Central processing unit, used to establish control algorithms; The central processing unit is also used to input the state vector input in parallel to two processing channels of the preset control algorithm, and output motion control commands for the logistics robot. An optimization learning unit is used to update the state vector based on the execution result of the action control command and generate learning signal parameters; The optimization learning unit is also used to set the optimization learning strategy of the control algorithm according to the learning signal parameters; The central processing unit includes: The first control module is used to acquire historical environmental data; Multiple environmental characteristic indicators are selected based on the historical environmental data; A scene semantic model and multiple environmental scenarios are established based on all the aforementioned environmental feature indicators; Establish the expected action commands for each environmental scenario in sequence; The expected action command includes: expected linear velocity and expected angular velocity; Construct a brain decision-making network based on all expected action instructions and the scene semantic model; Establish brain connectivity channels that link to the brain's decision-making network; Select multiple main equipment indicators; A threshold activation function and a weight table are established based on all the aforementioned device indicators; Construct a cerebellar adaptive network based on the threshold activation function and the weight table; Establish a cerebellar compensatory pathway that connects to the cerebellum adaptive network.

[0017] In some embodiments of this application, the central processing unit further includes: The first control module is used to acquire visual and point cloud data based on the state vector; The visual and point cloud data are input into the brain connection channel, and the scene semantic model extracts global scene semantics from the visual and point cloud data. Output basic action instructions based on the global scene semantics; The ontology perception signal is obtained based on the state vector; The proprioceptive signal is input into the cerebellar compensation channel and converted into a sparse pulse sequence through the threshold activation function. The sparse pulse sequence is mapped to the weight table through a hash mapping input space, and an instantaneous compensation action is generated. An environmental fluctuation quantity is generated based on the environmental state data. The weighting coefficient of the instantaneous compensation action is set according to the amount of environmental fluctuation; Based on the weighting coefficients, the basic action command and the instantaneous compensation action are vector-synthesized; The motion control command is generated based on the synthesis result.

[0018] In some embodiments of this application, the optimization learning unit includes: A signal acquisition module is used to establish a time axis, which includes multiple consecutive moments. Generate the expected environmental state for the next moment at the current moment; When the logistics robot finishes executing the action control command at the current moment, the sensing array acquires the actual observation state at the next moment; The sensory prediction error is calculated based on the expected environmental state and the actual observed state. Generate prediction error signal; Obtain the current task completion rate and calculate the sparse reward based on the task completion rate; Generate task reward signals; Generate learning signal parameters based on the prediction error signal and the task reward signal; An optimization module is used to build the reward feedback channel of the cerebellum adaptive network; Obtain the prediction error signal from the learned signal parameters; The prediction error signal is input into the reward feedback channel to generate the weight update parameters of the cerebellum adaptive network; Obtain the task reward signal from the learning signal parameters; Based on the task reward signal, determine whether to generate a migration update instruction for the brain decision network; The optimization learning algorithm is set based on the weight update parameters and migration update instructions.

[0019] Compared with existing technologies, the advantages of the logistics robot control method and system based on online transfer learning proposed in this application are as follows: By constructing a dual-pathway architecture that parallels a brain-based decision-making network (high-level strategy) and a cerebellum-based adaptive network (low-level compensation), functional decoupling of high-level decision-making and rapid adaptation is achieved. The brain-based decision-making network outputs basic motor commands based on cross-environment task planning and strategy transfer; the cerebellum-based adaptive network focuses on handling dynamic disturbances in logistics scenarios (such as changes in cargo weight and ground friction coefficient), providing real-time fine-tuning. This ensures the macro-optimal nature of the strategy and the system's rapid response to sudden changes.

[0020] By monitoring changes in the distribution of input features in real time, the system dynamically determines the update strategies for each layer of the brain's decision-making network. Through lightweight fine-tuning, it can quickly adapt to new scenarios, improve the efficiency of adapting to different environments, prevent catastrophic forgetting, and maintain the memory of learned environments. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a preferred embodiment of a logistics robot control method based on online transfer learning. Detailed Implementation

[0022] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but are not intended to limit the scope of this application.

[0023] In the description of this application, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0024] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0025] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0026] like Figure 1 As shown, a preferred embodiment of this application provides a logistics robot control method based on online transfer learning, comprising: S101: Collect environmental state data through the sensing array of the logistics robot, and generate a state vector based on the environmental state data; S102: Input the state vector in parallel to two processing channels of the preset control algorithm, and output the motion control command of the logistics robot; S103: Update the state vector based on the execution result of the motion control command, generate learning signal parameters, and set the optimization learning strategy of the control algorithm based on the learning signal parameters; The control algorithm includes: Brain decision-making network and cerebellum adaptive network; The two processing pathways include: the brain decision-making pathway and the cerebellum compensation pathway.

[0027] Specifically, the sensing array includes a visual channel, a point cloud channel, and a body perception channel. The visual channel (RGB-D camera) acquires image data of goods, shelves, and obstacles at a frequency of 10-30Hz. The point cloud channel (LiDAR) scans and acquires environmental geometric data at a frequency of 10-20Hz. The body perception channel acquires high-frequency data from joint encoders, motor current / torque sensors, and chassis IMU at a frequency of 200-1000Hz.

[0028] Specifically, the multi-mode fusion interface is used to uniformly process heterogeneous sensing information from multiple sources, such as lidar point clouds, RGB-D images, IMU inertial data, and motor current / torque, collected by the sensing array, to achieve spatiotemporal alignment and feature complementarity, thereby generating a real-time state vector.

[0029] Specifically, the state vector includes the current visual and point cloud data (generated by processing data collected from the visual and point cloud channels) and the body perception signal (generated by processing data collected from the body perception channel). The body perception signal includes joint angles, angular velocities, torque errors, etc.

[0030] Specifically, the Brain Decision Network is a deep neural network trained on the Transformer architecture. It is loaded with high-level semantic representations and macro-decision strategies of the learning environment. This network can be pre-trained through an online transfer learning framework and has meta-learning capabilities across environments.

[0031] Specifically, the cerebellar adaptive network is a lightweight network trained on a spiking neural network. It is responsible for high-frequency proprioceptive signal processing and real-time motion compensation, outputs fine-tuning torque, and corrects the execution deviation of the brain's decision-making network in dynamic environments.

[0032] Specifically, a preset control algorithm is established, including: Obtain historical environmental data; Multiple environmental characteristic indicators were selected based on historical environmental data; Establish a scene semantic model and multiple environmental scenarios based on all environmental feature indicators; Establish the expected action commands for each environmental scenario in sequence; The expected motion commands include: expected linear velocity and expected angular velocity; Construct a brain decision-making network based on all expected action instructions and scene semantic models; Establish brain connectivity channels that link to the brain's decision-making network; Select multiple main equipment indicators; Establish a threshold activation function and weight table based on all the device indicators; A cerebellar adaptive network was constructed based on a threshold activation function and a weight table. Establish a cerebellar compensatory pathway that connects to the cerebellum adaptive network.

[0033] Specifically, historical environmental data refers to data related to the working environment of logistics robots (such as shelf / rack layout, aisle width, shelf area, obstacle type, cargo height, cargo center distribution, ground friction coefficient, slope, and other parameters that affect the control motion decisions of logistics robots). Historical environmental data can be generated by filtering data recorded by logistics robots in historical tasks, providing data support for the initial training of the robot's decision-making network. This historical environmental data is then input into the Transformer architecture to extract various environmental feature indicators (such as shelf layout, aisle width, and cargo position). Based on the different real-time parameters corresponding to these environmental feature indicators, various environmental scenarios are constructed. By analyzing historical data, the optimal motion control parameters (i.e., expected motion commands) for the logistics robot in each environmental scenario are generated.

[0034] Specifically, the scene semantic model is trained on a Transformer-based encoder and can extract global scene semantics (i.e., real-time parameters of various environmental feature indicators) from visual and point cloud data through a self-attention mechanism. This facilitates the brain's decision network in matching corresponding action instructions.

[0035] Specifically, by analyzing the historical task record data of logistics robots, multiple body equipment indicators are selected, including but not limited to parameters that map the operating status of logistics robots such as joint angles, angular velocities, and torque errors.

[0036] It is understandable that, in the above embodiments, a dual-pathway architecture is constructed that parallels the brain-based decision-making network (high-level strategy) and the cerebellum-based adaptive network (low-level compensation) to achieve functional decoupling between high-level decision-making and rapid adaptation. The brain-based decision-making network outputs basic motor commands based on cross-environment task planning and strategy transfer; the cerebellum-based adaptive network focuses on handling dynamic disturbances in logistics scenarios (such as changes in cargo weight and ground friction coefficient), providing real-time fine-tuning. This ensures the macro-optimal nature of the strategy and the system's rapid response capability to sudden changes.

[0037] In a preferred embodiment of this application, the output motion control commands for the logistics robot include: Obtaining visual and point cloud data based on state vectors; Visual and point cloud data are input into the brain's connection channel, and the scene semantic model extracts global scene semantics from the visual and point cloud data. Output basic action instructions based on the global scene semantics; Acquiring ontology perception signals based on state vectors; Proprioceptive signals are input into the cerebellar compensation channel and converted into a sparse pulse sequence through a threshold activation function. The sparse pulse sequence is mapped to the weight table through a hash mapping input space, and instantaneous compensation actions are generated. Motion control commands are generated based on basic motion commands and instantaneous compensation commands.

[0038] Specifically, the scene semantic model extracts global scene semantics through a self-attention mechanism, generates real-time parameters for various environmental feature indicators, and dynamically adjusts the batch normalization parameters and some convolution weights of the feature extraction layer to achieve cross-environment feature alignment. The adapted features are then input into the brain decision network to output basic action commands (including the target linear velocity and target angular velocity of the logistics robot).

[0039] Specifically, the proprioceptive signals (joint angle, angular velocity, torque error) are input into the Granule Layer, converted into a sparse pulse sequence through a threshold activation function, with the activation rate controlled below 5%. The sparse pulse sequence is then mapped from the input space to a weight table through a hash mapping, and the weight values ​​of the activated units are queried and aggregated to generate instantaneous compensation actions.

[0040] Specifically, the cerebellar compensation channel operates at a frequency of 200Hz with a latency of less than 5ms.

[0041] Specifically, generating motion control commands includes: Environmental fluctuation data are generated based on environmental status data. The weighting coefficients for instantaneous compensation actions are set based on the amount of environmental fluctuations. Vector synthesis is performed on the basic motion command and the instantaneous compensation motion based on the weight coefficient; Generate motion control commands based on the synthesis results.

[0042] Specifically, basic motions and instantaneous compensation motions are generated by vector synthesis through basic motion instructions to generate control instructions.

[0043] Specifically, the synthesis process is as follows: A total =A brain +α.A cerebellum ; Where α is the dynamic weighting coefficient, calculated in real time by the environmental uncertainty estimation module: when a high dynamic disturbance (such as ground slippage or sudden load change) is detected, the cerebellar compensation weight is automatically increased. total For the synthesized action, A brain Basic action, A cerebellum For instantaneous compensation action; composite action A total After safety constraint checks (speed limit, acceleration smoothing), the signal is sent to the underlying motor driver via the CAN bus for execution.

[0044] Understandably, in the above embodiments, the brain-based decision network outputs basic movement commands based on cross-environment task planning and strategy transfer; the cerebellum-based adaptive network focuses on handling dynamic disturbances in the logistics scenario (such as changes in cargo weight and ground friction coefficient), providing real-time fine-tuning. This ensures the macro-optimal nature of the strategy and the system's rapid response to sudden changes.

[0045] In a preferred embodiment of this application, the generation of learning signal parameters includes: Create a timeline that includes multiple consecutive moments. Generate the expected environmental state for the next moment at the current moment; When the logistics robot finishes executing the action control command for the current moment, the sensing array acquires the actual observed state for the next moment. Calculate the sensory prediction error based on the expected environmental conditions and the actual observed conditions; Generate prediction error signal; Get the current task completion rate and calculate the sparse reward based on the task completion rate; Generate task reward signals; The learning signal parameters are generated based on the prediction error signal and the task reward signal.

[0046] Specifically, the sensory prediction error is calculated by comparing the predicted state s^t+1 based on a simplified internal model (such as a forward kinematics model) with the actual observed state st+1: et=st+1 s^t+1; This error signal directly reflects dynamic changes in the environment (such as pose deviation caused by tire slippage), serving as an instantaneous learning signal for the cerebellar adaptive network.

[0047] Specifically, sparse rewards rt are calculated based on task completion rates (such as cargo handling success rate, path length, and energy consumption) and used to periodically optimize the brain's decision-making strategies.

[0048] Specifically, the optimization learning strategy for the control algorithm is defined, including: Constructing a reward feedback channel for a cerebellar adaptive network; Obtain the prediction error signal from the learned signal parameters; The prediction error signal is input into the reward feedback channel to generate the weight update parameters for the cerebellum adaptive network. Obtain the task reward signal from the learning signal parameters; Determine whether to generate a transfer and update instruction for the brain's decision-making network based on the task reward signal; The optimization learning algorithm is set based on the weight update parameters and migration update instructions.

[0049] Specifically, the cerebellar adaptive network constructs a reward-independent... A rapid adaptive pathway with excitation feedback. Once the prediction error et is generated, it immediately drives the cerebellar weight update: Δw = ηfast et / K Where ηfast is the fast learning rate (usually set to 0.01-0.1), and K is the number of currently activated CMAC units. This update involves only local weight adjustments, has extremely low computational overhead, and can be completed within milliseconds after the action is executed, achieving rapid perturbation suppression similar to biological reflexes.

[0050] Specifically, by calculating the difference between the expected state and the actual state in real time, the cerebellar adaptive network is provided with richer and more timely learning guidance than traditional reward signals. This design enables the system to recognize changes in environmental features within milliseconds, just like the biological cerebellum, achieving precise online self-calibration. It is particularly suitable for high-precision operations such as fine-tuning of forklift fork height and obstacle avoidance in narrow passages.

[0051] Specifically, determining whether to generate transfer update instructions for the brain's decision-making network includes: Generate cumulative reward values ​​based on task reward signals; If the accumulated reward value exceeds the preset reward value threshold, a migration update instruction will be generated. The migration update instructions include: Based on the collected environmental interaction data, the distribution changes of environmental features are generated; The migration and update range of the brain's decision-making network is selected based on the distribution of changes in environmental characteristics; Perform lightweight adjustments based on the migration update scope.

[0052] Specifically, after the migration update instruction is completed, the sparse rewards calculated after each task are completed are accumulated to generate a cumulative reward value.

[0053] Specifically, the reward threshold can be set based on historical parameters. When the cumulative reward value is greater than the preset reward threshold, it indicates that the brain decision network is poorly adapted to the current environmental state of the logistics robot, and the brain decision network needs to be optimized according to the migration update instruction.

[0054] Specifically, environmental interaction data refers to the relevant data collected by the logistics robot in the current working environment. The changes in the distribution of environmental features are monitored in real time through maximum mean difference (MMD) or domain classifiers. The general feature extraction layer (low-level convolution) is frozen, and only the task-specific layer (high-level fully connected) is fine-tuned. If necessary, an adapter layer is introduced for feature alignment. Elastic weight consolidation (EWC) or progressive neural networks are used to prevent catastrophic forgetting during the transfer process and to maintain the knowledge memory of the learned environment.

[0055] Specifically, the strategy performance is continuously optimized by periodically performing lightweight fine-tuning.

[0056] It is understandable that in the above embodiments, by monitoring changes in the distribution of input features in real time, the update strategy of each layer of the brain's decision network is dynamically determined. Through lightweight fine-tuning, new scenarios can be quickly adapted, improving the efficiency of adaptation to different environments, while preventing catastrophic forgetting and maintaining the memory of the learned environment.

[0057] In another preferred embodiment of the logistics robot control method based on online transfer learning, based on any of the above preferred embodiments, a logistics robot control system based on online transfer learning is provided, comprising: The data processing unit is used to collect environmental state data through the sensing array of the logistics robot and generate a state vector based on the environmental state data. Central processing unit, used to establish control algorithms; The central processing unit is also used to input the state vector input in parallel to two processing channels of the preset control algorithm, and output motion control commands for the logistics robot. An optimized learning unit is used to update the state vector based on the execution result of the motion control command and generate learning signal parameters; The optimization learning unit is also used to set the optimization learning strategy of the control algorithm based on the learning signal parameters; The central processing unit includes: The first control module is used to acquire historical environmental data; Multiple environmental characteristic indicators were selected based on historical environmental data; Establish a scene semantic model and multiple environmental scenarios based on all environmental feature indicators; Establish the expected action commands for each environmental scenario in sequence; The expected motion commands include: expected linear velocity and expected angular velocity; Construct a brain decision-making network based on all expected action instructions and scene semantic models; Establish brain connectivity channels that link to the brain's decision-making network; Select multiple main equipment indicators; Establish a threshold activation function and weight table based on all the device indicators; A cerebellar adaptive network was constructed based on a threshold activation function and a weight table. Establish a cerebellar compensatory pathway that connects to the cerebellum adaptive network.

[0058] In a preferred embodiment of this application, the central processing unit further includes: The first control module is used to acquire visual and point cloud data based on the state vector; Visual and point cloud data are input into the brain's connection channel, and the scene semantic model extracts global scene semantics from the visual and point cloud data. Output basic action instructions based on the global scene semantics; Acquiring ontology perception signals based on state vectors; Proprioceptive signals are input into the cerebellar compensation channel and converted into a sparse pulse sequence through a threshold activation function. The sparse pulse sequence is mapped to the weight table through a hash mapping input space, and instantaneous compensation actions are generated. Environmental fluctuation data are generated based on environmental status data. The weighting coefficients for instantaneous compensation actions are set based on the amount of environmental fluctuations. Vector synthesis is performed on the basic motion command and the instantaneous compensation motion based on the weight coefficient; Generate motion control commands based on the synthesis results.

[0059] In a preferred embodiment of this application, the optimization learning unit includes: The signal acquisition module is used to establish a time axis, which includes multiple consecutive moments. Generate the expected environmental state for the next moment at the current moment; When the logistics robot finishes executing the action control command for the current moment, the sensing array acquires the actual observed state for the next moment. Calculate the sensory prediction error based on the expected environmental conditions and the actual observed conditions; Generate prediction error signal; Get the current task completion rate and calculate the sparse reward based on the task completion rate; Generate task reward signals; The learning signal parameters are generated based on the prediction error signal and the task reward signal; An optimization module is used to build the reward feedback channel of the cerebellum adaptive network; Obtain the prediction error signal from the learned signal parameters; The prediction error signal is input into the reward feedback channel to generate the weight update parameters for the cerebellum adaptive network. Obtain the task reward signal from the learning signal parameters; Determine whether to generate a transfer and update instruction for the brain's decision-making network based on the task reward signal; The optimization learning algorithm is set based on the weight update parameters and migration update instructions.

[0060] Based on the first concept of this application, a dual-pathway architecture is constructed, in parallel with a brain-based decision-making network (high-level strategy) and a cerebellum-based adaptive network (low-level compensation), to achieve functional decoupling between high-level decision-making and rapid adaptation. The brain-based decision-making network outputs basic motor commands based on cross-environment task planning and strategy transfer; the cerebellum-based adaptive network focuses on handling dynamic disturbances in logistics scenarios (such as changes in cargo weight and ground friction coefficient), providing real-time fine-tuning. This ensures the macro-optimal nature of the strategy and the system's rapid response to sudden changes.

[0061] According to the second concept of this application, by monitoring changes in the distribution of input features in real time, the update strategy of each layer of the brain's decision network is dynamically determined. Through lightweight fine-tuning, it can quickly adapt to new scenarios, improve the efficiency of adaptation to different environments, prevent catastrophic forgetting, and maintain the memory of the learned environment.

[0062] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and substitutions can be made without departing from the technical principles of this application, and these improvements and substitutions should also be considered within the scope of protection of this application.

Claims

1. A control method for logistics robots based on online transfer learning, characterized in that, include: The environmental state data is collected by the sensing array of the logistics robot, and a state vector is generated based on the environmental state data. The state vector input is fed in parallel to two processing channels of a preset control algorithm, and the motion control commands for the logistics robot are output. The state vector is updated based on the execution result of the action control command, and learning signal parameters are generated. The optimization learning strategy of the control algorithm is set based on the learning signal parameters. The control algorithm includes: Brain decision-making network and cerebellum adaptive network; The two processing channels include: the brain decision-making channel and the cerebellum compensation channel; The preset control algorithm includes: Obtain historical environmental data; Multiple environmental characteristic indicators are selected based on the historical environmental data; A scene semantic model and multiple environmental scenarios are established based on all the aforementioned environmental feature indicators; Establish the expected action commands for each environmental scenario in sequence; The expected action command includes: expected linear velocity and expected angular velocity; The brain decision-making network is constructed based on all expected action instructions and the scene semantic model; Establish the brain connection channel that connects to the brain decision-making network; Select multiple main equipment indicators; A threshold activation function and a weight table are established based on all the aforementioned device indicators; The cerebellum adaptive network is constructed based on the threshold activation function and the weight table. Establish the cerebellar compensation channel that is connected to the cerebellar adaptive network.

2. The logistics robot control method based on online transfer learning as described in claim 1, characterized in that, The output motion control commands for the logistics robot include: Visual and point cloud data are obtained based on the state vector; The visual and point cloud data are input into the brain connection channel, and the scene semantic model extracts global scene semantics from the visual and point cloud data. Output basic action instructions based on the global scene semantics; The ontology perception signal is obtained based on the state vector; The proprioceptive signal is input into the cerebellar compensation channel and converted into a sparse pulse sequence through the threshold activation function. The sparse pulse sequence is mapped to the weight table through a hash mapping input space, and an instantaneous compensation action is generated. The motion control command is generated based on the basic motion command and the instantaneous compensation motion.

3. The logistics robot control method based on online transfer learning as described in claim 2, characterized in that, Generating the motion control command includes: An environmental fluctuation quantity is generated based on the environmental state data. The weighting coefficient of the instantaneous compensation action is set according to the amount of environmental fluctuation; Based on the weighting coefficients, the basic action command and the instantaneous compensation action are vector-synthesized; The motion control command is generated based on the synthesis result.

4. The logistics robot control method based on online transfer learning as described in claim 3, characterized in that, The generated learning signal parameters include: Establish a timeline that includes multiple consecutive moments; Generate the expected environmental state for the next moment at the current moment; When the logistics robot finishes executing the action control command at the current moment, the sensing array acquires the actual observation state at the next moment; The sensory prediction error is calculated based on the expected environmental state and the actual observed state. Generate prediction error signal; Obtain the current task completion rate and calculate the sparse reward based on the task completion rate; Generate task reward signals; Learning signal parameters are generated based on the prediction error signal and the task reward signal.

5. The logistics robot control method based on online transfer learning as described in claim 4, characterized in that, The optimization learning strategy of the set control algorithm includes: Constructing a reward feedback channel for a cerebellar adaptive network; Obtain the prediction error signal from the learned signal parameters; The prediction error signal is input into the reward feedback channel to generate the weight update parameters of the cerebellum adaptive network; Obtain the task reward signal from the learning signal parameters; Based on the task reward signal, determine whether to generate a migration update instruction for the brain decision network; The optimization learning algorithm is set based on the weight update parameters and migration update instructions.

6. The logistics robot control method based on online transfer learning as described in claim 5, characterized in that, The determination of whether to generate a transfer update instruction for the brain's decision-making network includes: Generate a cumulative reward value based on the task reward signal; If the accumulated reward value exceeds the preset reward value threshold, a migration update instruction will be generated. The migration update instructions include: Based on the collected environmental interaction data, the distribution changes of environmental features are generated; The migration and update range of the brain decision network is selected based on the distribution of changes in the environmental characteristics. Perform lightweight adjustments based on the migration update scope.

7. A logistics robot control system based on online transfer learning, employing the logistics robot control method based on online transfer learning as described in any one of claims 1-6, characterized in that, include: The data processing unit is used to collect environmental state data through the sensing array of the logistics robot and generate a state vector based on the environmental state data. Central processing unit, used to establish control algorithms; The central processing unit is also used to input the state vector input in parallel to two processing channels of the preset control algorithm, and output motion control commands for the logistics robot. An optimization learning unit is used to update the state vector based on the execution result of the action control command and generate learning signal parameters; The optimization learning unit is also used to set the optimization learning strategy of the control algorithm according to the learning signal parameters; The central processing unit includes: The first control module is used to acquire historical environmental data; Multiple environmental characteristic indicators are selected based on the historical environmental data; A scene semantic model and multiple environmental scenarios are established based on all the aforementioned environmental feature indicators; Establish the expected action commands for each environmental scenario in sequence; The expected action command includes: expected linear velocity and expected angular velocity; Construct a brain decision-making network based on all expected action instructions and the scene semantic model; Establish brain connectivity channels that link to the brain's decision-making network; Select multiple main equipment indicators; A threshold activation function and a weight table are established based on all the aforementioned device indicators; Construct a cerebellar adaptive network based on the threshold activation function and the weight table; Establish a cerebellar compensatory pathway that connects to the cerebellum adaptive network.

8. The logistics robot control system based on online transfer learning as described in claim 7, characterized in that, The central processing unit further includes: The first control module is used to acquire visual and point cloud data based on the state vector; The visual and point cloud data are input into the brain connection channel, and the scene semantic model extracts global scene semantics from the visual and point cloud data. Output basic action instructions based on the global scene semantics; The ontology perception signal is obtained based on the state vector; The proprioceptive signal is input into the cerebellar compensation channel and converted into a sparse pulse sequence through the threshold activation function. The sparse pulse sequence is mapped to the weight table through a hash mapping input space, and an instantaneous compensation action is generated. An environmental fluctuation quantity is generated based on the environmental state data. The weighting coefficient of the instantaneous compensation action is set according to the amount of environmental fluctuation; Based on the weighting coefficients, the basic action command and the instantaneous compensation action are vector-synthesized; The motion control command is generated based on the synthesis result.

9. The logistics robot control system based on online transfer learning as described in claim 8, characterized in that, The optimization learning unit includes: A signal acquisition module is used to establish a time axis, which includes multiple consecutive moments. Generate the expected environmental state for the next moment at the current moment; When the logistics robot finishes executing the action control command at the current moment, the sensing array acquires the actual observation state at the next moment; The sensory prediction error is calculated based on the expected environmental state and the actual observed state. Generate prediction error signal; Obtain the current task completion rate and calculate the sparse reward based on the task completion rate; Generate task reward signals; Generate learning signal parameters based on the prediction error signal and the task reward signal; An optimization module is used to build the reward feedback channel of the cerebellum adaptive network; Obtain the prediction error signal from the learned signal parameters; The prediction error signal is input into the reward feedback channel to generate the weight update parameters of the cerebellum adaptive network; Obtain the task reward signal from the learning signal parameters; Based on the task reward signal, determine whether to generate a migration update instruction for the brain decision network; The optimization learning algorithm is set based on the weight update parameters and migration update instructions.

Images