A Deep Learning-Based Intelligent Service Robot Control System

Abstract

This invention relates to the field of intelligent robot control technology and provides a deep learning-based intelligent service robot control system. The system constructs a multimodal semantic reasoning model to fuse and model multi-source perception information such as vision, LiDAR, and radar. Within a unified semantic space, it performs tasks such as target detection, navigable area understanding, and dynamic obstacle analysis, improving the consistency of environmental perception. To address the training instability caused by non-stationary changes in the indoor environment, a smooth function-based two-layer optimization method is introduced to achieve adaptive optimization of multi-task trade-offs and fusion strategies, enhancing the stability and robustness of semantic reasoning. Furthermore, at the decision-making and control level, a hierarchical reinforcement learning method based on an enhanced Option-Critic architecture is employed to achieve synergy between task decision-making and continuous control, thereby improving the robot's safety and execution efficiency in complex service tasks.

Description

Technical Field

[0001] This invention relates to the field of intelligent robot control technology, and in particular to an intelligent service robot control system based on deep learning. Background Technology

[0002] With the development of artificial intelligence, computer vision, and robotics, intelligent technologies have been widely introduced into the field of indoor service robots to achieve environmental perception, task scheduling, and autonomous motion control. Existing intelligent service robots have improved automation levels and service efficiency to some extent. However, existing intelligent technologies often focus on single perception tasks or local function optimization. They typically model the indoor environment based on single-modal or weakly fused multimodal perception methods and use deep learning models with static parameter configurations or reinforcement learning strategies with fixed structures for decision control. This makes it difficult to achieve stable and collaborative intelligent behavior in indoor service scenarios with dense crowds, frequently changing obstacles, diverse task types, and continuously evolving environmental states. Especially under conditions of multi-task parallel semantic understanding, multimodal information fusion, and long-term operation, existing intelligent systems generally rely on human experience to set multi-task weights and fusion strategies. They lack adaptability to non-stationary changes in the environment and data distribution, and are prone to problems such as fluctuating perception results, discontinuous decision-making, or decreased control stability. At the same time, existing intelligent decision-making technologies often adopt single-layer strategies or rule-driven methods, lacking hierarchical decoupling and collaborative optimization mechanisms between task-level intent and underlying continuous control, making it difficult to balance safety, efficiency, and robustness in the execution of complex tasks. Therefore, how to further improve the capabilities of indoor service robots in multimodal perception collaborative modeling, multi-task semantic reasoning stability, and intelligent decision-making and control in complex task scenarios based on existing intelligent technologies remains an urgent problem to be solved in this field. Summary of the Invention

[0003] To address the shortcomings of existing intelligent service robots in multimodal perception collaboration, poor stability of multi-task semantic reasoning, and difficulty in effectively connecting task decision-making and continuous control in complex indoor environments, this invention provides a deep learning-based intelligent service robot control system. This system constructs a unified multimodal semantic reasoning model to fuse and model multi-source perception information such as vision and LiDAR, and completes key tasks such as target detection, navigable area understanding, and dynamic obstacle analysis within the same framework, thereby improving the consistency of environmental understanding. During model training, a smooth function-based two-layer optimization method is introduced to adaptively optimize multi-task trade-off parameters and cross-modal fusion strategies, improving the model's training stability and inference robustness in non-stationary indoor scenarios. At the decision-making and control level, a hierarchical reinforcement learning method based on an enhanced Option-Critic architecture is adopted to effectively decouple and coordinate task-level decision-making with underlying continuous control, thereby improving the decision continuity, safety, and execution efficiency of the intelligent service robot in complex service scenarios.

[0004] This invention provides a deep learning-based intelligent service robot control system, which is integrated into an intelligent service robot in indoor service scenarios, such as hospital wards, elderly care facilities, hotel floors, government service halls, and airport terminals. The system includes: a robot, a multimodal perception and synchronization module, an edge reasoning module, a task parsing and scheduling module, an environment mapping module, a deep learning decision control module, and a motion execution module; wherein:

[0005] The robot includes a mobile chassis, power supply and management components, human-machine interaction components, and service execution components;

[0006] The multimodal perception and synchronization module is connected to the robot body and is used to collect environmental perception data and robot state data and perform time alignment. The multimodal perception and synchronization module uses hardware timestamp synchronization to synchronize multi-source data to obtain synchronized multimodal data. The synchronized multimodal data is then written to a shared cache queue for subsequent inference and control reading.

[0007] The edge inference module is located within the robot body and includes a CPU processing unit, a GPU acceleration unit, a storage unit, and a real-time communication bus. It performs preprocessing and feature encoding on synchronous multimodal data. Preprocessing includes image distortion correction, scale normalization, point cloud denoising and voxel downsampling, radar clutter suppression, and sensor extrinsic parameter alignment to obtain preprocessed multimodal data. A multimodal semantic inference model is initialized, and a smooth function-type two-layer optimization method is introduced to train the model, resulting in a trained multimodal semantic inference model. This trained model then performs semantic inference on the preprocessed multimodal data, outputting intermediate control decision results. These intermediate results include: target detection results, passable area semantic segmentation results, dynamic obstacle motion estimation results, service target localization results, and interaction object recognition confidence scores. The output of the inference model is presented in a unified manner. A semantic structure is written into the environment representation data structure and a query interface is provided to the task parsing and scheduling module and the deep learning decision control module. The construction process of the smooth function-type bilayer optimization method is as follows: Based on the function-type bilayer optimization framework, to address the problem of large noise in the outer layer supergradient estimation and unstable parameter updates in the non-stationary training environment of the multimodal semantic reasoning model, a smooth function-type bilayer optimization method is constructed by introducing a time-smoothing supergradient mechanism and combining it with the inner layer and the accompanying approximation error constraint. The construction process of the multimodal semantic reasoning model is as follows: Using BEVFusion as the backbone network, five task prediction heads are set in parallel at the output of the backbone network. The five task prediction heads are the target detection head, the passable area semantic segmentation head, the dynamic obstacle motion estimation head, the service target localization head, and the interactive object recognition confidence head, thus constructing the multimodal semantic reasoning model.

[0008] The task parsing and scheduling module receives voice and touch task commands from the human-computer interaction component, performs semantic parsing and parameter extraction on the task commands, and forms a task description that includes the target location, service type, delivery method, and time constraints. During the generation of the task description, the module calls the query interface provided by the edge inference module to obtain environmental representation information related to the current task. Based on the task description and environmental representation information, the task description is decomposed into a sequence of sub-tasks, including navigation arrival, passage through channels, obstacle avoidance and vehicle meeting, interaction confirmation, capture and execution, return to home, and recharging docking. Success conditions, failure rollback strategies, and safety constraint levels are configured for each sub-task, and the sub-task scheduling results are obtained.

[0009] The environmental mapping module is used to achieve localization, mapping, and local accessibility representation in indoor environments. This module is based on laser SLAM, combining wheel speed odometry and IMU data to estimate the robot's current pose and output positioning confidence. On this basis, a topological map representing indoor structural relationships is constructed, and a local cost map is generated within the robot's local area. The local cost map integrates the static obstacle occupancy probability, dynamic obstacle predicted trajectory, accessible area boundary, narrow passage constraint, and safety distance expansion layer. Based on the local cost map and the topological constraints extracted from the topological map, the module extracts accessible areas, generates passage skeletons, and filters access constraints to generate candidate access corridors to guide subsequent task scheduling and control decisions.

[0010] The deep learning decision control module includes a high-level policy network and a low-level control policy network. The high-level policy network, based on the Option-Critic architecture, introduces a dynamic option creation mechanism based on the return variance trend. This optimizes the Option-Critic architecture's ability to cover complex state spaces, the complementarity of option functions, and its structural adaptability to policy learning stability during non-stationary training, under a fixed number of options, thus constructing an enhanced Option-Critic model. The system utilizes the currently activated subtasks and their corresponding task-level objectives and execution constraints, local cost maps, and candidate pathways from the subtask scheduling results. The system takes the intermediate results of control decisions as input and outputs the operation sub-objective through the enhanced Option-Critic model. The low-level control strategy network takes the operation sub-objective and the robot's current state as input and outputs the chassis control quantity and the robot arm control quantity through the SAC continuous control strategy network. The chassis control quantity includes linear velocity, angular velocity and velocity-steering angle control quantity, and the robot arm control quantity includes joint increment, end-effector pose increment and impedance parameters. Based on the operation sub-objective, the system combines the local cost map and candidate passage corridors to perform trajectory unfolding and parameterization on the operation sub-objective, and further generates a trajectory point sequence for the motion control and execution interface module to track and execute.

[0011] The motion execution module includes a chassis controller, a robotic arm controller, and an end effector controller. The chassis controller uses chassis control quantities as desired motion commands and combines them with trajectory point sequences to execute a combination of velocity and attitude closed-loop control. The velocity closed loop constrains the tracking errors of the chassis linear velocity and angular velocity to the desired values, while the attitude closed loop suppresses heading and lateral deviations during path tracking. Based on the closed-loop control results, the chassis controller generates torque or speed commands and sends corresponding drive control signals to the motor driver via the EtherCAT bus. The robotic arm controller uses robotic arm control quantities as desired control inputs and combines them with the description of the robotic arm's motion intention by the operation sub-objective to execute joint space trajectory tracking control. During the control process, the robotic arm controller performs online verification of joint limits, velocity constraints, and acceleration constraints for each joint and limits the control commands when a constraint violation risk is detected. The end effector controller controls the opening and closing of the gripper, the lifting and lowering of the pallet, and the actions of the pick-and-place mechanism. It is linked with the interactive confirmation process and completes the retrieval action after the delivery confirmation conditions are met. Execution feedback is sent back to the deep learning decision control module for online correction and risk monitoring.

[0012] Furthermore, the multimodal semantic reasoning model is initialized, and a smooth function-type two-layer optimization method is introduced to train the multimodal semantic reasoning model, resulting in a trained multimodal semantic reasoning model. The trained multimodal semantic reasoning model is then used to perform semantic reasoning on the preprocessed multimodal data, outputting intermediate results for control decisions. This process specifically includes the following steps:

[0013] Step S1: Initialize the multimodal semantic reasoning model based on the preset multimodal fusion network structure. After the network structure initialization is completed, perform unified initialization on the network parameters of each prediction head, and further configure the set of adjustable outer parameters corresponding to each prediction head. The set of adjustable outer parameters includes multi-task loss weight parameters, cross-modal fusion weight parameters, and time consistency constraint weight parameters.

[0014] Step S2: Construct a two-layer optimization training structure suitable for non-stationary training environments. The training process of the multimodal semantic reasoning model is divided into an inner optimization process and an outer optimization process. The inner optimization process is used to construct the inner objective function, and the outer optimization process is used to construct the outer objective function, so as to formalize the model training process into a functional two-layer optimization problem suitable for non-stationary environments.

[0015] Step S3: Under the condition of fixing the set of adjustable parameters of the outer layer, the inner layer objective function is used as the optimization objective. Based on the data distribution corresponding to the current time round, the network weight parameters of the multimodal semantic reasoning model are trained and updated to minimize the training loss of multimodal semantic reasoning, thereby obtaining the inner layer optimized model parameters corresponding to the current time round.

[0016] Step S4: After completing the inner layer optimization and obtaining the inner layer optimized model parameters, the outer layer optimization process takes the state of the inner layer optimized model parameters as input to construct a stochastic hypergradient estimate of the outer layer objective function with respect to the set of outer layer adjustable parameters. This estimate is used to characterize the update direction of the outer layer adjustable parameters. The stochastic hypergradient estimate includes explicit gradient terms and implicit gradient terms. By jointly constructing the explicit and implicit gradient terms, the stochastic hypergradient estimate result of the outer layer in the current time round is obtained, providing gradient information for subsequent time-smooth outer layer parameter updates. The stochastic hypergradient estimate includes: 1. Explicit gradient terms, which represent the direct gradient components of the outer layer objective function with respect to the set of outer layer adjustable parameters. These terms characterize the direct impact of the set of outer layer adjustable parameters on the performance evaluation index of multimodal semantic reasoning. 2. Implicit gradient terms, which represent the gradient components introduced through the inner layer optimal conditions. These terms characterize the influence of the set of outer layer adjustable parameters on the multimodal semantic reasoning results by affecting the inner layer optimization process and the model training trajectory.

[0017] Step S5: To address the issues of high noise and oscillations in single-round outer stochastic hypergradient estimation in non-stationary training environments, in the current training round, the time-smoothing hypergradient mechanism is used to perform a weighted average of the multi-round outer stochastic hypergradient estimation results obtained from multiple training rounds within a preset time window according to a preset time weight coefficient, thereby constructing the corresponding time-smoothing hypergradient; based on the time-smoothing hypergradient, gradient descent is performed to update the set of adjustable outer parameters.

[0018] Step S6: After completing the gradient descent update of the outer adjustable parameter set, introduce convergence stability constraints on the inner optimization approximation error and the approximation error of the adjoint variable to evaluate the error propagation and cumulative effect of the outer stochastic hypergradient estimator constructed based on the time smoothing hypergradient mechanism; after completing the above error evaluation, determine the upper bound of the cumulative error of the outer stochastic hypergradient estimator in the current training round, determine the stability and robustness of the outer adjustable parameter update process under the condition of approximate calculation, and use it as the constraint basis for the continuous update of the outer adjustable parameter set in subsequent training rounds;

[0019] Step S7: Repeat steps S3 to S6 until the update magnitude of the outer adjustable parameter set meets the convergence condition, then training terminates, and the parameters of the multimodal semantic reasoning model optimized by the smooth function-type two-layer optimization method are solidified to obtain the trained multimodal semantic reasoning model. Semantic reasoning is then performed on the preprocessed multimodal data through the trained multimodal semantic reasoning model to output intermediate results of control decisions.

[0020] Furthermore, the process of outputting operational sub-objectives through the enhanced Option-Critic model specifically includes the following steps:

[0021] Step B1: Construct a joint state input representation for high-level policy decision-making based on the currently active subtask and its corresponding task-level objectives and execution constraints, local cost map and candidate passageways, and intermediate results of control decisions in the subtask scheduling results;

[0022] Step B2: Input the joint state input representation into the high-level policy network based on the Option-Critic architecture; select the current execution option from the existing option set maintained in the high-level policy network in the current state through the option policy network, and determine the continued execution or termination switching condition of the current execution option in combination with the option termination function to generate the initial operation sub-objective. The operation sub-objective is used to describe the intermediate navigation target, interactive station target or operation intention target in the current stage.

[0023] Step B3: During training, perform statistical analysis on the cumulative returns of each option in the existing option set across multiple evaluation rounds, calculate the return variance of each option at different time rounds; construct a variance evolution sequence for the return variance over time, and perform integral and trend fitting processing on the variance evolution sequence to obtain a return variance trend index that characterizes the stability of the option in the current training stage.

[0024] Step B4: Compare the return variance trend index with a preset threshold. When it is detected that the return variance trend of a certain option is continuously higher than the threshold during the continuous training phase, and it is determined that the existing option set cannot stably cover part of the current state space, the dynamic option creation mechanism is triggered to generate new options to expand the option set size of the high-level policy network.

[0025] Step B5: After creating a new option, perform bias initialization on the option value function of the new option to suppress its indiscriminate exploration of the global state space in the initial stage;

[0026] Step B6: Call the historical interaction trajectory data generated during training to perform experience replay pre-training on the new option; during the replay process, only the state-action samples that are more likely to be activated by the new option under the current option selection mechanism are updated, and the other samples do not participate in the training of the new option; in the above way, the new option has the initial coverage of the blind spot of the state space before resuming online training;

[0027] Step B7: After completing the initialization and pre-training of the new options, update the option set of the high-level policy network; re-execute the option selection and termination decision in the current state, and output the corresponding operation sub-objective for subsequent trajectory unfolding and low-level control policy network use.

[0028] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0029] First, it effectively improves the consistency of perception and the reliability of environmental understanding in complex indoor environments.

[0030] This invention constructs a multimodal semantic reasoning model based on BEV fusion, unifying multi-source perception information from vision, LiDAR, and radar into a consistent spatial semantic representation. Within the same model framework, it performs tasks such as target detection, navigable area understanding, dynamic obstacle motion estimation, and service object identification in parallel, achieving information sharing and collaborative reasoning among multiple tasks. Compared to existing technologies that model multiple tasks independently with fragmented results, this invention avoids interference from semantic inconsistencies between different perception tasks, significantly improving the robot's overall perception integrity and spatial semantic understanding accuracy in complex indoor environments. This provides a more reliable environmental foundation for subsequent task scheduling and motion control.

[0031] Second, it significantly enhances the training stability and long-term robustness of multimodal semantic reasoning models in non-stationary scenarios.

[0032] This invention introduces a smooth function-based two-layer optimization method during model training. It jointly optimizes multi-task trade-offs, cross-modal fusion strategies, and time consistency constraints as adjustable outer parameters. Furthermore, it suppresses the impact of random noise and environmental changes on parameter updates through a time-smooth hypergradient update mechanism, thereby reducing performance fluctuations of the multi-task model under non-stationary training conditions. Through this mechanism, this invention achieves stable output of the multimodal semantic reasoning model across time and scenarios, effectively solving the problem of reasoning result fluctuations caused by fixed manual parameters or unstable hyperparameter updates in existing intelligent technologies. This enhances the reliability and maintainability of the robot during long-term continuous service.

[0033] Third, it improves the continuity of decision-making, security, and overall execution efficiency under complex service tasks.

[0034] This invention constructs a hierarchical reinforcement learning decision-making model based on an enhanced Option-Critic architecture. A dynamic option generation mechanism based on reward variance trends is introduced at the high-level layer, enabling the high-level policy network to adaptively expand the option space, stably cover complex task state regions, and effectively decouple task-level decision-making from the underlying continuous control. Combined with the low-level continuous control policy network, coordinated control of chassis movement and operation execution is achieved. Therefore, this invention reduces frequent policy switching and manual rule intervention in complex indoor service scenarios, enabling robots to efficiently complete multi-stage service tasks while meeting safety constraints, significantly improving overall decision continuity, execution reliability, and service efficiency. Attached Figure Description

[0035] Figure 1This is a trajectory diagram of the evolution of the adjustable outer layer parameters proposed in Example 2;

[0036] Figure 2 This is a flowchart illustrating the enhanced Option-Critic model proposed in Example 4. Detailed Implementation

[0037] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0038] Example 1: This invention provides a deep learning-based intelligent service robot control system. This system is integrated into an intelligent service robot in indoor service scenarios, such as hospital wards, elderly care facilities, hotel floors, government service halls, and airport terminals. The system includes: a robot, a multimodal perception and synchronization module, an edge reasoning module, a task parsing and scheduling module, an environment mapping module, a deep learning decision control module, and a motion execution module; wherein:

[0039] The robot comprises a mobile chassis, a power supply and management component, a human-machine interface component, and a service execution component. The mobile chassis is a differential wheel type, equipped with a drive motor, wheel speed encoder, and motor driver. The service execution component includes a robotic arm and an end effector for performing grasping, delivery, delivery confirmation, or item retrieval actions. The power supply and management component includes a power battery, a battery management system (BMS), and a power distribution module for supplying power to the computing unit, sensors, and actuators, and outputting operating status information such as battery level, health status, and temperature. The human-machine interface component includes a touch screen, a microphone array, and a speaker for receiving voice / touch commands and outputting voice / graphic interactive feedback.

[0040] The multimodal perception and synchronization module is connected to the robot body to collect environmental perception data and robot state data and perform time alignment. The environmental perception data includes RGB images, depth maps, 3D laser point clouds, and millimeter-wave radar echoes. The robot state data includes IMU inertial measurement data, wheel speed and odometer readings, joint angles / joint torques, collision edge signals, and battery state data. The multimodal perception and synchronization module uses hardware timestamp synchronization to synchronize multimodal data and writes the synchronized multimodal data into a shared cache queue for subsequent inference and control reading.

[0041] The edge inference module is located within the robot body and includes a CPU processing unit, a GPU acceleration unit, a storage unit, and a real-time communication bus. It performs preprocessing and feature encoding on synchronous multimodal data. Preprocessing includes image distortion correction, scale normalization, point cloud denoising and voxel downsampling, radar clutter suppression, and sensor extrinsic parameter alignment to obtain preprocessed multimodal data. A multimodal semantic inference model is initialized, and a smooth function-type two-layer optimization method is introduced to train the model, resulting in a trained multimodal semantic inference model. This trained model then performs semantic inference on the preprocessed multimodal data, outputting intermediate control decision results. These intermediate results include: target detection results, passable area semantic segmentation results, dynamic obstacle motion estimation results, service target localization results, and interaction object recognition confidence scores. The output of the inference model is presented in a unified manner. A semantic structure is written into the environment representation data structure and a query interface is provided to the task parsing and scheduling module and the deep learning decision control module. The construction process of the smooth function-type bilayer optimization method is as follows: Based on the function-type bilayer optimization framework, to address the problem of large noise in the outer layer supergradient estimation and unstable parameter updates in the non-stationary training environment of the multimodal semantic reasoning model, a smooth function-type bilayer optimization method is constructed by introducing a time-smoothing supergradient mechanism and combining it with the inner layer and the accompanying approximation error constraint. The construction process of the multimodal semantic reasoning model is as follows: Using BEVFusion as the backbone network, five task prediction heads are set in parallel at the output of the backbone network. The five task prediction heads are the target detection head, the passable area semantic segmentation head, the dynamic obstacle motion estimation head, the service target localization head, and the interactive object recognition confidence head, thus constructing the multimodal semantic reasoning model.

[0042] The task parsing and scheduling module receives voice and touch task commands from the human-computer interaction component, performs semantic parsing and parameter extraction on the task commands, and forms a task description that includes the target location, service type, delivery method, and time constraints. During the generation of the task description, the module calls the query interface provided by the edge inference module to obtain environmental representation information related to the current task. Based on the task description and environmental representation information, the task description is decomposed into a sequence of sub-tasks, including navigation arrival, passage through channels, obstacle avoidance and vehicle meeting, interaction confirmation, capture and execution, return to home, and recharging docking. Success conditions, failure rollback strategies, and safety constraint levels are configured for each sub-task, and the sub-task scheduling results are obtained.

[0043] The environmental mapping module is used to achieve localization, mapping, and local accessibility representation in indoor environments. This module is based on laser SLAM, combining wheel speed odometry and IMU data to estimate the robot's current pose and output positioning confidence. On this basis, a topological map representing indoor structural relationships is constructed, and a local cost map is generated within the robot's local area. The local cost map integrates the static obstacle occupancy probability, dynamic obstacle predicted trajectory, accessible area boundary, narrow passage constraint, and safety distance expansion layer. Based on the local cost map and the topological constraints extracted from the topological map, the module extracts accessible areas, generates passage skeletons, and filters access constraints to generate candidate access corridors to guide subsequent task scheduling and control decisions.

[0044] The deep learning decision control module includes a high-level policy network and a low-level control policy network. The high-level policy network, based on the Option-Critic architecture, introduces a dynamic option creation mechanism based on the return variance trend. This optimizes the Option-Critic architecture's ability to cover complex state spaces, the complementarity of option functions, and its structural adaptability to policy learning stability during non-stationary training, under a fixed number of options, thus constructing an enhanced Option-Critic model. The system utilizes the currently activated subtasks and their corresponding task-level objectives and execution constraints, local cost maps, and candidate pathways from the subtask scheduling results. The system takes the intermediate results of control decisions as input and outputs the operation sub-objective through the enhanced Option-Critic model. The low-level control strategy network takes the operation sub-objective and the robot's current state as input and outputs the chassis control quantity and the robot arm control quantity through the SAC continuous control strategy network. The chassis control quantity includes linear velocity, angular velocity and velocity-steering angle control quantity, and the robot arm control quantity includes joint increment, end-effector pose increment and impedance parameters. Based on the operation sub-objective, the system combines the local cost map and candidate passage corridors to perform trajectory unfolding and parameterization on the operation sub-objective, and further generates a trajectory point sequence for the motion control and execution interface module to track and execute.

[0045] In this embodiment, after the robot receives the task instruction to "deliver medication to ward 305 and complete delivery confirmation":

[0046] Operational sub-objective: Proceed along the right side of the corridor to the door of ward 305, and decelerate, align, and stand in a position 0.6 meters to the right front of the bed;

[0047] Chassis control parameters: linear velocity: 0.45 m / s; angular velocity: 0.12 rad / s; switching to: target speed: 0.30 m / s, steering angle: 6° in narrow areas;

[0048] Robotic arm control parameters: Joint increments: 2nd joint: +4.5°, 4th joint: -3.0°; Impedance parameters: Stiffness = 300 N / m, Damping = 35 Ns / m.

[0049] The motion execution module includes a chassis controller, a robotic arm controller, and an end effector controller. The chassis controller uses chassis control inputs as desired motion commands and, combined with a trajectory point sequence, executes a combination of speed and attitude closed-loop control. The speed closed-loop is used to constrain the tracking errors of the chassis's linear and angular velocities relative to the desired values, while the attitude closed-loop is used to suppress heading and lateral deviations during path tracking. Based on the closed-loop control results, the chassis controller generates torque or speed commands and sends corresponding drive control signals to the motor driver via the EtherCAT bus. The robotic arm controller uses robotic arm control inputs as desired control inputs. The system takes input and combines it with the description of the robotic arm's motion intention based on the operation sub-objectives to execute joint space trajectory tracking control. During the control process, the robotic arm controller performs online verification of the joint limits, speed constraints, and acceleration constraints of each joint, and limits the control commands when a constraint violation risk is detected. The end effector controller controls the opening and closing of the gripper, the lifting and lowering of the pallet, and the actions of the pick-and-place mechanism, and links with the interactive confirmation process to complete the retrieval action after the delivery confirmation conditions are met. Execution feedback (such as wheel speed, posture error, joint error, and torque anomaly) is sent back to the deep learning decision control module for online correction and risk monitoring.

[0050] Example 2, according to Figure 1 This embodiment is based on Embodiment 1. In this embodiment, a multimodal semantic reasoning model is initialized, and a smooth function-type two-layer optimization method is introduced to train the multimodal semantic reasoning model to obtain a trained multimodal semantic reasoning model. The process of performing semantic reasoning on preprocessed multimodal data and outputting intermediate results of control decisions through the trained multimodal semantic reasoning model specifically includes the following steps:

[0051] Step S1: Initialize the multimodal semantic reasoning model based on the preset multimodal fusion network structure. After the network structure initialization is completed, perform unified initialization on the network parameters of each prediction head, and further configure the set of adjustable outer parameters corresponding to each prediction head. The set of adjustable outer parameters includes multi-task loss weight parameters, cross-modal fusion weight parameters, and time consistency constraint weight parameters.

[0052] generate Figure 1 : Evolution trajectory diagram of adjustable parameters in the outer layer; Figure 1In the figure, the horizontal axis represents the training round t, and the vertical axis represents the adjustable parameters (weights) of the outer layer. The figure shows the evolution trajectory of the multi-task loss weight parameters, cross-modal fusion weight parameters, and time consistency constraint weight parameters of the multimodal semantic reasoning model during non-stationary training. There are three environmental change points in the figure, namely environmental change point 1, environmental change point 2, and environmental change point 3.

[0053] Step S2: Construct a two-layer optimization training structure suitable for non-stationary training environments. The training process of the multimodal semantic reasoning model is divided into an inner optimization process and an outer optimization process. The inner optimization process is used to construct the inner objective function, and the outer optimization process is used to construct the outer objective function, so as to formalize the model training process into a functional two-layer optimization problem suitable for non-stationary environments.

[0054] Step S3: Inner Layer Optimization Process. Under the condition of a fixed set of adjustable outer layer parameters, the inner layer objective function is used as the optimization objective. Based on the data distribution corresponding to the current time epoch, the network weight parameters of the multimodal semantic reasoning model are trained and updated to minimize the multimodal semantic reasoning training loss, thereby obtaining the inner layer optimized model parameters corresponding to the current time epoch. The inner layer objective function is expressed as:

[0055] ;

[0056] in, This represents the time epoch index during the training process. This represents the set of control parameters for adjusting the multi-task trade-offs and fusion strategies of the multimodal semantic reasoning model. In this invention, these parameters include: 1. Multi-task loss weight parameters (e.g., weights for detection / segmentation / motion estimation / localization / interaction recognition); 2. Cross-modal fusion weight parameters (e.g., the proportion of different modal features in BEV fusion, gating coefficients, etc.); 3. Temporal consistency constraint weight parameters (used to constrain output stability across rounds and environmental changes). This represents the inner model function corresponding to the multimodal semantic reasoning model. The inner model function is composed of the BEVFusion backbone network and its various task prediction heads, and its parameters correspond to the network weight parameters of the multimodal semantic reasoning model. Indicates the first Each time cycle, outer adjustable parameters Under fixed conditions, the optimal inner-layer model function obtained through the inner-layer optimization process is the parameter state of the inner-layer optimization model in the current time cycle. Indicates the inner model function As optimization variables, the optimal inner model function is solved by minimizing the objective function; Indicates the first The distribution of multimodal training data corresponding to each time round is used to reflect the differences in data distribution caused by changes in the environment, scenario and task conditions of the indoor service robot at different time stages, thereby reflecting the non-stationarity of the training environment; This represents multimodal input data samples, including environmental perception data such as RGB images, depth maps, laser point clouds, and millimeter-wave radar echoes, as well as necessary time synchronization or pose assistance information; Indicates input sample Corresponding supervision and annotation information; This represents the multimodal semantic reasoning training loss function used in the inner layer optimization process, used to measure the performance given the adjustable parameters of the outer layer. Under the condition, inner model function For input samples The generated prediction results With real labeling Deviation between;

[0057] Step S4: After completing the inner layer optimization and obtaining the inner layer optimization model parameters, the outer layer optimization process takes the state of the inner layer optimization model parameters as input to construct a stochastic hypergradient estimate of the outer layer objective function with respect to the set of outer layer adjustable parameters, which is used to characterize the update direction of the outer layer adjustable parameters. The stochastic hypergradient estimate includes explicit gradient terms and implicit gradient terms. By jointly constructing the explicit gradient terms and implicit gradient terms, the outer layer stochastic hypergradient estimate result under the current time cycle is obtained, which provides gradient information for subsequent time-smooth outer layer parameter updates.

[0058] Stochastic hypergradient estimation includes: 1. Explicit gradient terms, which represent the direct gradient components of the outer objective function with respect to the set of outer adjustable parameters, used to characterize the direct impact of the set of outer adjustable parameters on the performance evaluation index of multimodal semantic reasoning; 2. Implicit gradient terms, which represent the gradient components introduced through the inner optimal conditions, used to characterize the indirect impact of the set of outer adjustable parameters on the results of multimodal semantic reasoning by influencing the inner optimization process and the model training trajectory.

[0059] The outer objective function is expressed as:

[0060] ;

[0061] in, Indicates the adjustable parameters of the outer layer As optimization variables, the update direction and magnitude of the outer parameters are determined by minimizing the outer objective function; Represents the outer objective function. This represents the comprehensive evaluation loss function used for outer layer optimization, which characterizes the performance of a given set of adjustable outer layer parameters. Under the condition of inner layer optimization model Input samples in non-stationary multimodal sensing environments The semantic reasoning output relative to the supervised information Deviations in consistency, robustness, and decision availability across time-phased rounds;

[0062] The stochastic hypergradient estimate of the outer adjustable parameter is expressed as:

[0063] ;

[0064] in, Indicates the first The set of adjustable outer parameters for each round This represents stochastic hypergradient estimation. This represents the direct partial derivative of the outer objective function with respect to the outer adjustable parameters, corresponding to the explicit gradient term; This represents the mixed partial derivative of the inner objective function with respect to the outer adjustable parameters and the model output, corresponding to the implicit gradient term; The adjoint function is used to characterize the response of the inner optimal solution to changes in the adjustable parameters of the outer layer. It is used to map the indirect influence of the inner optimization process on the outer layer parameters to the outer gradient space. Indicates the first In each time cycle, with a fixed outer adjustable parameter Under the given conditions, the approximate optimal solution of the inner model function corresponding to the inner optimization model parameters is obtained after training and updating the multimodal semantic reasoning model through the inner optimization process; This indicates that the inner model is based on the input samples. The model output results generated above;

[0065] Step S5: To address the issues of high noise and oscillations in single-round outer-layer stochastic supergradient estimation in non-stationary training environments, in the current training round, a time-smoothing supergradient mechanism is used to perform a weighted average of the multi-round outer-layer stochastic supergradient estimation results obtained from multiple training rounds within a preset time window, according to a preset time weight coefficient, to construct the corresponding time-smoothing supergradient. Based on the time-smoothing supergradient, gradient descent is performed to update the set of adjustable outer-layer parameters, enabling the multimodal semantic reasoning model to adaptively adjust the multi-task trade-offs and cross-modal feature fusion strategies during training across time rounds, thereby continuously adapting to the perception changes and decision-making needs in non-stationary indoor service scenarios. While maintaining unbiasedness of the true outer-layer gradient within the window in the expected sense, the time-smoothing supergradient effectively suppresses high-frequency noise introduced by random sampling and environmental non-stationarity, thereby reducing the supergradient estimation variance and improving the stability and robustness of the outer-layer parameter update process.

[0066] Step S6: After updating the outer adjustable parameter set using gradient descent, convergence and stability constraints are introduced for the inner optimization approximation error and the adjoint variable approximation error. The error propagation and cumulative effects of the outer stochastic hypergradient estimator built based on the time-smoothing hypergradient mechanism are evaluated. Specifically, the cumulative impact of the inner optimization approximation error over time is characterized to assess the impact on the accuracy of the outer hypergradient estimation when the inner model parameters have not reached the exact optimal solution. Similarly, the cumulative impact of the adjoint variable approximation error over time is characterized to assess the error introduced by the approximate solution of the adjoint variable on the implicit gradient term estimation. After completing the above error evaluation, the upper bound of the cumulative error of the outer stochastic hypergradient estimator in the current training epoch is determined. This determines the stability and robustness of the outer adjustable parameter update process under approximate calculation conditions and serves as a constraint for the continuous updating of the outer adjustable parameter set in subsequent training epochs.

[0067] The upper bound of the cumulative error of the outer stochastic hypergradient estimator satisfies the following constraint:

[0068] ;

[0069] in, Indicates the total number of training rounds. This represents the time smoothing scale parameter used to construct the time-smoothed outer stochastic hypergradient, characterizing the temporal range of historical hypergradient estimates involved in the smoothing calculation in the current training epoch and their equivalent weight strength. Indicates the first The time-smoothed outer stochastic hypergradient estimate obtained during round training; Indicates the first The gradient of the true outer layer objective function corresponding to each round of training; Represents the vector norm; , and A positive constant coefficient is used to unify error terms from different sources onto the same scale; This represents the upper bound of the variance of the outer stochastic hypergradient estimate. Represents the variance term of random noise. This represents the cumulative term of the inner layer approximation error. This represents the cumulative term of approximation error of the accompanying variable;

[0070] Step S7: Repeat steps S3 to S6 until the update magnitude of the outer adjustable parameter set meets the convergence condition, then training terminates, and the parameters of the multimodal semantic reasoning model optimized by the smooth function-type two-layer optimization method are solidified to obtain the trained multimodal semantic reasoning model. Semantic reasoning is then performed on the preprocessed multimodal data through the trained multimodal semantic reasoning model to output intermediate results of control decisions.

[0071] Example 3: This example is based on Example 1. In this example, a multimodal semantic reasoning model is initialized, and stochastic gradient descent (SGD) is introduced to train the multimodal semantic reasoning model to obtain a trained multimodal semantic reasoning model. Semantic reasoning is performed on preprocessed multimodal data through the trained multimodal semantic reasoning model to output intermediate results of control decisions.

[0072] Example 4, according to Figure 2 This embodiment is based on Embodiment 2. In this embodiment, the process of outputting the operation sub-objective through the enhanced Option-Critic model specifically includes the following steps:

[0073] Step B1: State Construction: Based on the currently active subtask and its corresponding task-level objectives and execution constraints, local cost map and candidate passageways in the subtask scheduling results, and intermediate results of control decisions, construct a joint state input representation for high-level policy decisions;

[0074] Step B2: Option Selection: Input the joint state input representation into the high-level policy network based on the Option-Critic architecture; select the current execution option from the existing option set maintained in the high-level policy network in the current state through the option policy network, and determine the continued execution or termination switching conditions of the current execution option in combination with the option termination function to generate the initial operation sub-objective. The operation sub-objective is used to describe the intermediate navigation target, interactive station target or operation intention target in the current stage.

[0075] Step B3: Variance Evaluation: During training, statistical analysis is performed on the cumulative returns of each option in the existing option set across multiple evaluation rounds to calculate the return variance of each option at different time rounds; a variance evolution sequence is constructed on the return variance over time, and integral and trend fitting processing is performed on the variance evolution sequence to obtain a return variance trend index that characterizes the stability of the option in the current training stage.

[0076] Step B4: Option Expansion: Compare the return variance trend index with a preset threshold. When it is detected that the return variance trend of a certain option is continuously higher than the threshold during the continuous training phase, it is determined that the existing option set cannot stably cover part of the current state space. Then, the dynamic option creation mechanism is triggered to generate new options to expand the option set size of the high-level policy network.

[0077] Step B5: Parameter initialization: After creating a new option, perform bias initialization on the option value function of the new option to suppress its indiscriminate exploration of the global state space in the initial stage;

[0078] Step B6: Replay Pre-training: Call the historical interaction trajectory data generated during training to perform experience replay pre-training on the new option; during the replay process, only the state-action samples that are more likely to be activated by the new option under the current option selection mechanism are updated, and the other samples do not participate in the training of the new option; in the above way, the new option has the initial coverage of the blind spot of the state space before resuming online training;

[0079] Step B7: Sub-target output: After completing the initialization and pre-training of new options, update the option set of the high-level policy network; re-execute option selection and termination decision in the current state, and output the corresponding operation sub-target for subsequent trajectory unfolding and low-level control policy network use.

[0080] In conventional techniques, the Option-Critic architecture is used to output operational sub-objectives, and the process specifically includes the following steps:

[0081] Step C1: Based on the currently active subtask and its corresponding task-level objectives and execution constraints, local cost map and candidate passageways, and intermediate results of control decisions in the subtask scheduling results, construct a joint state input representation for high-level policy decisions;

[0082] Step C2: Input the joint state input representation into a high-level policy network based on the Option-Critic architecture, and pre-set a fixed-size set of options during the training initialization phase;

[0083] Step C3: Using the option policy network, select the currently executing option from the fixed option set in the current state, and determine the conditions for continuing execution or terminating the currently executing option based on the corresponding option termination function;

[0084] Step C4: Generate the corresponding operation sub-target based on the current execution options.

[0085] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A deep learning-based intelligent service robot control system, characterized in that, The system includes: The multimodal perception and synchronization module collects environmental perception data and robot state data and performs time alignment to obtain synchronized multimodal data; The edge reasoning module performs preprocessing and feature encoding on synchronous multimodal data to obtain preprocessed multimodal data; it initializes the multimodal semantic reasoning model, introduces a smooth function-type two-layer optimization method to train the multimodal semantic reasoning model, and obtains a trained multimodal semantic reasoning model. The trained multimodal semantic reasoning model is then used to perform semantic reasoning on the preprocessed multimodal data to output intermediate results of control decisions. The task parsing and scheduling module receives task instructions and obtains subtask scheduling results; The environmental mapping module generates a local cost map and candidate passageways. The deep learning decision control module includes a high-level policy network and a low-level control policy network: the high-level policy network takes the currently active subtask and its corresponding task-level objectives and execution constraints, local cost maps and candidate passageways, and intermediate control decision results as inputs from the subtask scheduling results, and outputs the operation sub-objective using an enhanced Option-Critic model; the low-level control policy network takes the operation sub-objective and the robot's current state as inputs, and outputs the chassis control quantity and the robotic arm control quantity through the SAC continuous control policy network. The motion execution module receives the output from the deep learning decision control module and executes the corresponding motion control.

2. The intelligent service robot control system based on deep learning according to claim 1, characterized in that: The task understanding and scheduling module receives task instructions and performs semantic parsing and parameter extraction on the task instructions to form a task description. During the process of generating the task description, it calls the query interface provided by the edge inference module to obtain environmental representation information. Based on the task description and environmental representation information, the task description is decomposed into a sequence of sub-tasks, and success conditions, failure rollback strategies and security constraint levels are configured to obtain the sub-task scheduling results.

3. The intelligent service robot control system based on deep learning according to claim 1, characterized in that: The construction process of the multimodal semantic reasoning model is as follows: using BEVFusion as the backbone network, five task prediction heads are set in parallel at the output end of the backbone network, namely, the target detection head, the passable area semantic segmentation head, the dynamic obstacle motion estimation head, the service target localization head, and the interactive object recognition confidence head, to construct the multimodal semantic reasoning model.

4. The intelligent service robot control system based on deep learning according to claim 1, characterized in that: The construction process of the smooth function-type bilayer optimization method is as follows: Based on the functional bilayer optimization framework, to address the problems of high noise in the outer layer hypergradient estimation and unstable parameter updates in the non-stationary training environment of multimodal semantic reasoning models, a smooth function-type bilayer optimization method is constructed by introducing a time-smoothing hypergradient mechanism and combining it with inner layer and adjoint approximation error constraints.

5. The intelligent service robot control system based on deep learning according to claim 4, characterized in that: The process of initializing a multimodal semantic reasoning model, training the model using a smooth function-based two-layer optimization method, and outputting intermediate control decision results includes the following steps: Step S1: Perform unified initialization of the network parameters of each prediction head of the multimodal semantic reasoning model, and further configure the set of adjustable outer parameters corresponding to each prediction head; Step S2: Construct a two-layer optimization training structure, dividing the training process of the multimodal semantic reasoning model into an inner optimization process and an outer optimization process; wherein, the inner optimization process constructs the inner objective function, and the outer optimization process constructs the outer objective function; Step S3: Under the condition of fixing the set of adjustable parameters of the outer layer, the inner layer objective function is used as the optimization objective. Based on the data distribution corresponding to the current time round, the network weight parameters of the multimodal semantic reasoning model are trained and updated to obtain the inner layer optimized model parameters. Step S4: After completing the inner layer optimization and obtaining the inner layer optimization model parameters, the outer layer optimization process takes the state of the inner layer optimization model parameters as input to construct the stochastic hypergradient estimate of the outer layer objective function with respect to the set of outer layer adjustable parameters; the stochastic hypergradient estimate includes explicit gradient terms and implicit gradient terms. By jointly constructing the explicit gradient terms and implicit gradient terms, the outer layer stochastic hypergradient estimate result under the current time cycle is obtained. Step S5: Using a time-smoothing supergradient mechanism, the outer layer stochastic supergradient estimation results obtained from multiple training rounds within a preset time window are weighted and averaged according to a preset time weight coefficient to construct a time-smoothing supergradient; based on the time-smoothing supergradient, gradient descent is performed to update the set of adjustable outer layer parameters. Step S6: After completing the gradient descent update of the outer adjustable parameter set, introduce the convergence stability constraint of the inner optimization approximation error and the approximation error of the adjoint variable, evaluate the error propagation and cumulative effect of the outer stochastic supergradient estimator constructed based on the time smoothing supergradient mechanism, and use it as the constraint basis for the continuous update of the outer adjustable parameter set in subsequent training rounds. Step S7: Repeat steps S3 to S6 until the update magnitude of the outer adjustable parameter set meets the convergence condition, training terminates, the parameters of the multimodal semantic reasoning model are solidified, and the trained multimodal semantic reasoning model is obtained. Semantic reasoning is performed on the preprocessed multimodal data through the trained multimodal semantic reasoning model to output intermediate results of control decisions.

6. The intelligent service robot control system based on deep learning according to claim 5, characterized in that: The set of adjustable parameters in the outer layer includes multi-task loss weight parameters, cross-modal feature fusion weight parameters, and time consistency constraint weight parameters.

7. The intelligent service robot control system based on deep learning according to claim 5, characterized in that: Explicit gradient terms represent the direct gradient components of the outer objective function with respect to the set of outer adjustable parameters; implicit gradient terms represent the gradient components introduced by the inner optimal conditions, used to characterize how the set of outer adjustable parameters influences the inner optimization process and the model training trajectory.

8. The intelligent service robot control system based on deep learning according to claim 1, characterized in that: The enhanced Option-Critic model is constructed by introducing a dynamic Option creation mechanism based on the return variance trend on the basis of the Option-Critic architecture. This optimizes the Option-Critic architecture's ability to cover complex state spaces, the complementarity of option functions, and the structural adaptability of policy learning stability during non-stationary training under the condition of a fixed number of options, thus constructing the enhanced Option-Critic model.

Images