A control method and system for an autonomous mobile charging robot

Abstract

This invention discloses a control method and system for an autonomous mobile charging robot, belonging to the interdisciplinary field of mobile robotics and smart charging. The method includes: constructing an environmental map and locating the robot using multimodal sensors; navigating to the target area based on global path planning; during navigation, using a lightweight deep learning object detection network to detect charging interfaces and dynamic obstacles in real time, and outputting a visual perception reliability score; predicting the trajectory of the detected dynamic obstacles, generating a spatiotemporal correlation penalty term, and incorporating this penalty term as a constraint into the cost function of an improved TEB local path planner, while dynamically adjusting the weight coefficients in the cost function; controlling the robot's movement according to the optimized local path to achieve dynamic obstacle avoidance and docking with the charging interface. This invention achieves deep collaboration between perception and planning, improving the robustness and success rate of autonomous charging in complex environments.

Description

Technical Field

[0001] This invention relates to the intersection of mobile robots and smart charging, specifically to a control method and system for an autonomous mobile charging robot. Background Technology

[0002] With the rapid development of the electric vehicle industry, the number of electric vehicles in China continues to grow. However, the construction of charging stations is lagging behind, and the existing charging mode is mainly based on fixed charging stations, which has problems such as limited coverage and low flexibility. The battery swapping mode faces challenges such as high equipment costs and difficulties in battery standardization. "Charging difficulty" has become a key pain point restricting the promotion of electric vehicles.

[0003] Research on mobile charging robots has been conducted both domestically and internationally: PowerHydrant in the US developed an automatic charging device based on a robotic arm in 2014; Tesla launched a snake-like automatic charging robot in 2015; and Graz University of Technology in Austria, in collaboration with BMW and other companies, developed an automatic charging system in 2018. In China, FAW Hongqi produced a prototype of a mobile intelligent charging robot in 2022, and Zhongneng Congcong released a mobile charging device integrating energy storage and charging technology in 2024. However, existing technologies still have shortcomings: 1. Insufficient positioning and navigation accuracy: Traditional navigation technologies are prone to path deviations in complex indoor environments, with positioning errors typically exceeding 5cm, making it difficult to meet the high-precision requirements of charging interface docking; 2. Poor visual recognition stability: Most solutions rely on a single visual sensor, resulting in low accuracy in charging interface recognition under varying light intensity or occlusion conditions; 3. Low integration of the control system: Poor coordination between sensor data fusion, underlying drivers, and upper-level functional modules leads to high response delays, affecting charging efficiency; 4. Insufficient ease of operation: The lack of remote control interfaces suitable for ordinary users makes it difficult for some groups to quickly get started.

[0004] To address the aforementioned technical challenges, there is an urgent need for a control method and system for autonomous mobile charging robots that can significantly improve system stability, accuracy, and ease of use. Summary of the Invention

[0005] The purpose of this invention is to provide a control method and system for an autonomous mobile charging robot. Addressing the problems of sluggish dynamic obstacle avoidance response, poor robustness of charging interface recognition, and low system coordination efficiency in existing mobile charging robots under complex environments such as changing lighting and dynamic obstacle interference, this invention provides a control method and system that deeply coordinates multimodal perception and local planning, and dynamically adjusts planning weights based on visual reliability.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: On one hand, the present invention provides a control method for an autonomous mobile charging robot, comprising the following steps: Mapping and localization: Constructing an environmental map based on multimodal sensor data and locating the robot's pose in real time; Global path planning: After receiving the charging command, global path planning is performed based on the pre-built environment map to control the robot to move to the target area; Visual perception: Utilizing a lightweight deep learning object detection network deployed on edge computing units, the acquired images are processed in real time to detect charging ports and dynamic obstacles, and the visual perception reliability score and dynamic obstacle detection results are output. Trajectory prediction: Predict the trajectory of detected dynamic obstacles and generate predicted trajectory information; Local path planning: Based on the robot's current pose, global path, LiDAR data, the dynamic obstacle detection results, and predicted trajectory information, local path planning is performed using an improved TEB algorithm; wherein, the cost function of the improved TEB algorithm includes a spatiotemporal correlation penalty term generated based on the predicted trajectory information, and the planning parameters are dynamically adjusted according to the output results of the visual perception step, and the planning parameters include at least the weight coefficient of the dynamic obstacle distance term. Docking and charging: When the robot arrives at the target area, the robot's actuator is driven to complete docking and charging based on the charging interface detection results.

[0007] Furthermore, the visual perception reliability score is calculated as follows: the input image is classified into scenes, the scene type and light environment quality score are output, the light environment quality score is multiplied by the category confidence score output by the target detection network, and the comprehensive reliability score is obtained after time-series smoothing filtering.

[0008] Furthermore, the trajectory prediction of the dynamic obstacle uses a Kalman filter to estimate the motion state of the dynamic obstacle based on the detection results of multiple consecutive frames, and extrapolates the predicted trajectory point sequence in the future time period.

[0009] Furthermore, the spatiotemporal correlation penalty term is: in, For spatiotemporal correlation penalty terms, The preset dynamic obstacle weights, For safe distance threshold, The robot pose output by the TEB planner. For the predicted dynamic obstacle trajectory points, and Robot pose and trajectory points The corresponding timestamp, The attenuation coefficient is... This reflects the characteristic that forecast uncertainty increases over time.

[0010] Furthermore, dynamically adjusting the planning parameters based on the output of the visual perception step includes at least one of the following methods: (a) Sensor Trust Level Adaptive Allocation: Based on the confidence interval of the visual perception reliability score, the trust level allocation ratio of the visual sensor, LiDAR and inertial measurement unit is adjusted in stages; wherein, when the visual perception reliability score is in the high confidence interval, the visual sensor obtains the dominant trust level; when the visual perception reliability score is in the low confidence interval, the trust level of the visual sensor is suppressed and the LiDAR obtains the dominant trust level. (b) Planning weight switch adjustment: When a dynamic obstacle is detected, increase the weight coefficient of the dynamic obstacle distance term and correspondingly decrease the weight coefficient of the path length term; when in a narrow passage, increase the weight coefficient of the speed change term; (c) Blind docking mode: When the visual perception reliability score is lower than the preset threshold, visual guidance is completely abandoned, and docking is completed based on historical docking records and lidar positioning.

[0011] Furthermore, the dynamic adjustment of planning parameters also includes continuous nonlinear adjustment based on scenario complexity: Scene complexity metrics Defined as: in, The number of dynamic targets detected. To calculate the spatial clustering degree of the target using kernel density estimation, Assess the reliability of visual perception. , , These are preset weighting coefficients; Adjust dynamic obstacle weights using the following formula And time optimality weight : in, This serves as the baseline value for the dynamic obstacle weight. This serves as the baseline value for the time-optimal weight. To adjust the amplitude coefficient, To reduce the coefficient, This is the saturation threshold.

[0012] Furthermore, the lightweight deep learning object detection network has at least one of the following features: End-to-end non-maximum suppression inference architecture: The network adopts a dual-head training mechanism, using both one-to-many and one-to-one supervision during the training phase, and retaining only the output of the one-to-one head during the inference phase, without the need for non-maximum suppression post-processing. Small target perception label allocation mechanism: Multiple anchor points are assigned to small targets at the charging interface for supervised training; Multi-scale linear attention enhancement module: Introduces multi-branch deep convolution in the backbone network to extract multi-scale features in parallel, and combines linear attention to capture global long-range dependencies; Multi-diffractor refiner: The neck network is divided into multiple heads along the channel dimension, and depthwise separable convolutions with different dilation rates are used for each head. Through channel interleaving and point convolution fusion, the refined extraction of multi-receptive field features is achieved. Multi-scene adaptive visual perception module: includes parallel scene classification branch and dynamic feature enhancement branch. Based on the scene classification results, it adaptively selects low light enhancement, backlight correction, dynamic blur compensation or high-resolution upsampling processing for small targets, and outputs a visual perception reliability score.

[0013] Furthermore, the lightweight deep learning object detection network, after channel pruning and INT8 quantization, is deployed on a Raspberry Pi edge computing platform and communicates data with zero copy via shared memory.

[0014] Furthermore, the cost function of the improved TEB algorithm includes a path length term, a static obstacle distance term, a dynamic obstacle distance term, an acceleration change term, and a velocity change term, wherein the dynamic obstacle distance term employs a spatiotemporal correlation penalty term based on the predicted trajectory, and the optimization objective of the cost function is to minimize the following expression: in, , , These are the weighting coefficients for the path length term, the static obstacle distance term, and the acceleration change term, respectively. , These are the adjusted dynamic obstacle weights and time-optimal weights, respectively. As a preset positive number, For path length, The distance from the robot to the static obstacle. For acceleration changes, This represents the change in speed.

[0015] On the other hand, the present invention provides an autonomous mobile charging robot control system for implementing the aforementioned autonomous mobile charging robot control method, comprising: The main control unit adopts a hierarchical architecture of upper computer and lower computer. The upper computer is used to run SLAM mapping, path planning and visual recognition algorithms, while the lower computer is used for low-level drive control. The multimodal sensor unit, including lidar, inertial measurement unit and monocular camera, is used to collect environmental perception data and robot status data; The drive execution unit, including a mobile chassis, a motor drive module, and a robotic arm, is used to drive the robot to move and complete the charging interface docking according to the instructions of the main control unit. The interaction unit includes a voice module, a display module, and a remote control module, which are used to receive user commands and provide feedback on the robot's status. The host computer is equipped with a lightweight deep learning object detection network that has undergone pruned quantization and runs an improved TEB algorithm.

[0016] The beneficial effects of the above scheme are as follows: First, in terms of positioning, navigation, and dynamic obstacle avoidance, the positioning error of this invention can be controlled within 2 cm, fully meeting the high-precision docking requirements of the charging interface; the obstacle detection response time is no more than 0.5 seconds. In particular, by coordinating the visual detection module with the improved local path planning in a closed loop, the robot can perceive and predict the movement trajectory of dynamic obstacles, further shortening the obstacle avoidance response time, and keeping the path deviation after obstacle avoidance within a small range, which is significantly better than traditional solutions.

[0017] Secondly, regarding visual recognition stability, this invention employs a lightweight, improved deep learning model that integrates multi-scale feature extraction and small-target optimization mechanisms. This achieves a recognition accuracy of over 95% for charging interfaces across a wide range of illumination variations, effectively solving the problem of recognition failures caused by illumination or occlusion in existing technologies. Through model pruning and quantization compression, the overall visual model size is reduced, improving real-time processing speed on embedded platforms and enabling high-precision, low-latency detection capabilities at the edge.

[0018] Third, in terms of adaptability to complex environments, this invention introduces a visual reliability-driven weight adaptive mechanism. When poor lighting or dynamic blurring leads to a decrease in visual recognition reliability, the system automatically increases the trust in other sensors and dynamically adjusts the weight coefficients of various costs in the path planner. In extreme cases, it even triggers a blind docking mode, thereby improving robustness and reliability in complex scenarios.

[0019] Fourth, in terms of overall system efficiency and ease of operation, the entire process from receiving a charging command to completing the docking is significantly faster than similar solutions. Users can initiate the charging process with a single click via a mobile application, view the charging progress in real time, and even stop the charging process in an emergency, greatly reducing the barrier to entry. Furthermore, the modular software framework design makes the system easily expandable with new sensors or algorithms, effectively reducing subsequent upgrade costs. Attached Figure Description

[0020] Figure 1 This is a functional design block diagram of the autonomous mobile charging robot of the present invention; Figure 2 This is a flowchart illustrating the overall workflow of the present invention; Figure 3 This is a flowchart of the host computer navigation control process of the present invention; Figure 4 This is a flowchart of the host computer visual recognition process of the present invention; Figure 5 This is a flowchart of the lower-level machine control process of the present invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0022] It should be noted that, unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The hardware, sensors, and software development environments employed are all commercially available or well-known in the art.

[0023] This invention provides an autonomous mobile charging robot control system and method. By deeply coordinating multimodal sensor fusion, a lightweight deep learning target detection network, and an improved TEB local path planning algorithm, it solves the problems of low positioning accuracy, unstable visual recognition, delayed obstacle avoidance response, and poor system coordination of existing mobile charging robots in complex lighting and dynamic obstacle environments. Example 1

[0024] like Figure 1 As shown, this embodiment provides a control system for an autonomous mobile charging robot. The system includes a main control unit, a multimodal sensor unit, a drive execution unit, an interaction unit, and a charging simulation unit. Each unit works collaboratively through standardized interfaces and communication protocols to jointly complete autonomous navigation and automatic charging tasks.

[0025] I. Main Control Unit The main control unit adopts a hierarchical architecture of "host computer-slave computer".

[0026] The host computer, based on a Raspberry Pi 4B (4GB RAM), runs Ubuntu 20.04 and the ROS Noetic framework. It is responsible for sensor data fusion, SLAM mapping, global and local path planning, visual recognition algorithm execution, and data interaction between modules. Specifically, the host computer runs gmapping for map building, move_base for navigation, an improved TEB algorithm for local path planning and dynamic obstacle avoidance, and deploys a lightweight deep learning object detection network with pruned and INT8 quantized branches for real-time recognition of charging ports and dynamic obstacles.

[0027] The lower-level controller uses an Arduino Mega2560 microcontroller to handle the underlying drive control, including motor speed control and robotic arm servo angle control. The lower-level controller communicates with the upper-level controller via serial port, receiving speed and servo commands from the upper-level controller, and feeding back encoder pulse data and motor drive status to the upper-level controller.

[0028] II. Multimodal sensor unit The multimodal sensor unit includes a lidar, an IMU module, and a monocular camera, used to collect environmental perception and robot status data.

[0029] LiDAR: The YDLIDAR X2 model is selected, based on the triangulation principle, with a ranging range of 0.12-8m, a scanning frequency of 5-8Hz, and a ranging frequency of 3000Hz. The LiDAR generates high-density point cloud data for map building, obstacle detection, and localization matching.

[0030] The IMU module, based on the MPU6050, integrates a three-axis accelerometer and a three-axis gyroscope, connecting to the Raspberry Pi via an I2C interface. IMU data is used to assist in correcting the robot's posture, particularly providing orientation reference in scenarios where LiDAR features are sparse.

[0031] Monocular camera: A USB-interface camera with a resolution of 320×240 pixels and a frame rate of 120 frames per second. The camera is calibrated with a 9×6 checkerboard pattern to eliminate lens distortion and is used to capture images of the charging port. During navigation, the camera is also used for visual detection of dynamic obstacles.

[0032] In addition, the system can be equipped with smoke and temperature sensors for real-time monitoring of the charging environment safety.

[0033] III. Drive Execution Unit The drive execution unit includes a mobile chassis, a motor drive module, and a robotic arm, which is used to execute motion commands and docking commands issued by the main control unit.

[0034] Mobile chassis: Employs four TT motors with AB-phase Hall encoders (reduction ratio 48:1, rated speed 100rpm). The chassis is equipped with Mecanum wheels, enabling omnidirectional movement with a maximum speed of 0.5m / s and a positioning accuracy of ±2cm. The encoders provide real-time feedback of the motor speed for closed-loop speed control and odometer calculation.

[0035] Motor drive module: L298N is selected, supporting 5-35V drive voltage. Each module can independently control two motors. The drive module adjusts the motor speed through PWM signal and achieves closed-loop speed control in conjunction with the encoder.

[0036] Robotic arm: Constructed from 3D-printed PLA material, equipped with three 180° servo motors to control chassis rotation, upper arm lifting, and lower arm extension / retraction. The end effector of the robotic arm houses a simulated charging gun and power cable. The servo motor control accuracy is ±1°. The robotic arm is used to precisely dock the charging interface after the robot reaches the target parking space.

[0037] IV. Interactive Unit The interactive unit includes a voice module, a display module, and a remote control module, which are used to realize human-computer interaction and command reception.

[0038] Voice module: Utilizes a voice dialogue development board based on the ESP32-C3 core, supporting natural Chinese voice interaction. Users can initiate charging, check status, or perform an emergency stop via voice commands.

[0039] Display module: Integrated OLED display (128×64 resolution) to display battery level, charging progress, fault information, and current robot status in real time. The display is mounted on the top of the robot for easy viewing by the user.

[0040] Remote control module: Employs an HC-04 Bluetooth module with a communication range of 10 meters. It connects to a mobile application, enabling one-button start charging, target parking space selection, charging progress monitoring, and emergency stop functions. Users can complete the entire charging process via the mobile application.

[0041] V. Charging Simulation Unit The charging simulation unit uses an Arduino Nano microcontroller, paired with an 18650 battery (7.4V / 2500mAh), a photosensitive sensor (to detect the insertion of the charging gun), an active buzzer (for fault alarms), and six LEDs (to simulate charging progress). When the charging gun is inserted, the LEDs light up in a sequential manner, and a "charging complete" signal is sent via Bluetooth when fully charged.

[0042] VI. System Software and Algorithm Implementation The system software is based on the ROS Noetic framework and adopts a layered design, including a perception layer, a data processing layer, a decision control layer, and an application layer. Each layer interacts with the topic through ROS nodes.

[0043] The perception layer processes multi-sensor data, including LiDAR data, IMU data, visual images, encoder data, and charging status data, through the Sensor_msgs package.

[0044] The data processing layer cleans, fuses, and extracts features from the raw data. Image preprocessing uses OpenCV for noise reduction and distortion correction, while point cloud processing uses the pcl_ros library to downsample the LiDAR point cloud (voxel filtering, voxel size 0.01m). 3 The data is segmented from the ground. Data fusion employs a spatiotemporal graph convolutional network to fuse IMU pose data and LiDAR point clouds, correcting robot pose errors. The fusion frequency is 50Hz. The visual recognition module extracts the red circular marker feature of the charging interface using a deep learning network and calculates the center coordinates and radius.

[0045] The decision control layer integrates a SLAM mapping module, an AMCL positioning module, an improved TEB path planning module, a visual recognition module, and a PID speed control module. The SLAM mapping module, based on the gmapping function package, combines LiDAR and IMU data to construct a two-dimensional grid map with a resolution of 0.05 m / pixel. The AMCL positioning module employs an adaptive Monte Carlo positioning algorithm with a positioning error ≤2 cm and an update frequency of 20 Hz.

[0046] The visual recognition module employs a lightweight deep learning object detection network. This network is based on the YOLOv26 architecture and features multi-scene adaptive improvements, including an end-to-end non-maximum suppression inference architecture, a small object perception label allocation mechanism, a multi-scale linear attention enhancement module, and a multi-expansion channel refiner. To enable edge deployment, the network undergoes channel pruning and INT8 quantization, compressing the model size to 2.3MB and achieving a forward inference speed of 25fps. The visual recognition module also incorporates a multi-scene adaptive perception front-end, automatically selecting and activating the low-light enhancement branch, backlight correction branch, dynamic blur compensation branch, or small object high-resolution upsampling branch based on the scene classification results of the input image.

[0047] The improved TEB algorithm is the core of the decision control layer. The TEB algorithm obtains a smooth and collision-free local path by optimizing the cost function. The cost function includes a path length term, a static obstacle distance term, a dynamic obstacle distance term, an acceleration change term, and a velocity change term. The dynamic obstacle distance term employs a spatiotemporal correlation penalty term based on the predicted trajectory, and its expression is: in, For spatiotemporal correlation penalty terms, The preset dynamic obstacle weights, For safe distance threshold, The robot pose output by the TEB planner. For the predicted dynamic obstacle trajectory points, and Robot pose and trajectory points The corresponding timestamp, The attenuation coefficient is... This reflects the characteristic that forecast uncertainty increases over time.

[0048] This penalty allows the robot to not only consider the current position of obstacles when optimizing its trajectory, but also to proactively avoid the space and time windows that they may occupy in the future.

[0049] The weight coefficients of the TEB algorithm are not fixed, but are dynamically adjusted in multiple levels based on the output results of the visual perception steps. The system first scores the visual perception reliability. The confidence level of the visual sensor, LiDAR, and IMU is adjusted according to the confidence interval. When When the value exceeds a preset high threshold, the visual sensor gains dominant trust; when... When the distance is below a preset low threshold, the LiDAR gains dominant trust. Secondly, when a dynamic obstacle is detected, the system increases the weight coefficient of the obstacle distance term and correspondingly decreases the weight coefficient of the path length term; when in a narrow passage, the weight coefficient of the speed change term is increased. Based on this, the system also defines a scene complexity index C, and accordingly performs tanh nonlinear adjustment on the dynamic obstacle weight and time optimality weight, achieving a smooth switch from an "efficiency-first" to a "conservative and safe" behavior mode in complex environments. When the threshold for blind docking is lower than the threshold, the system triggers blind docking mode, completely abandoning visual guidance and completing docking based on historical docking records and LiDAR positioning.

[0050] The PID speed control module implements PID control of the motor, with a proportional coefficient Kp=4.99, a derivative coefficient Kd=0.979, and an integral coefficient Ki=0.0005, achieving a speed control accuracy of ±0.01m / s. The obstacle avoidance control module, based on LiDAR data, triggers emergency braking and replans the path when an obstacle is detected at a distance ≤30cm, with a response time ≤0.5 seconds.

[0051] The application layer provides charging process control and remote control services. The charging process control module receives charging commands from the mobile application and sequentially initiates the mapping, navigation, identification, docking, charging, and return processes. The remote control module enables command interaction between the mobile application and the lower-level device via Bluetooth protocol.

[0052] VII. System Workflow like Figures 2-5 As shown, the complete process of the system from receiving the charging command to completing the charging is as follows.

[0053] (1) Command reception: The user selects the target parking space through the mobile APP, and the APP sends the command to the lower-level machine via Bluetooth. The lower-level machine forwards the command to the upper-level machine. (2) Mapping and navigation: The host computer starts gmapping for mapping and move_base for navigation. It plans the path by combining LiDAR and IMU data and moves the robot to the vicinity of the target parking space. During the navigation process, the vision recognition module detects dynamic obstacles in real time and outputs a reliability score. The improved TEB algorithm performs dynamic obstacle avoidance based on this.

[0054] (3) Visual recognition: After reaching the target parking space, the monocular camera and visual recognition module are activated to detect the charging interface: if it is not detected, the robotic arm chassis servo motor swings back and forth within a range of ±30°; if it is detected, the error is calculated and the servo motor is driven to adjust the angle, and it stops when the error is ≤2mm; (4) Simulated charging: After the charging gun is inserted into the simulated charging system, the photosensitive sensor triggers a signal, and the LED beads light up in a continuous process. (5) Task completion: After charging is completed, a "charging complete" signal is sent, the host computer drives the robotic arm to retract the charging gun, and the robot automatically returns to the standby point.

[0055] Throughout the process, all units work together to ensure the charging task is completed efficiently and safely. In mixed tests involving multiple scenarios including normal lighting, low lighting, strong backlight, and dynamic blur, the overall charging success rate of this system can reach up to 96.5%, the dynamic obstacle avoidance response time is shortened to less than 0.4s, and the path deviation after obstacle avoidance is ≤1.2cm, verifying the effectiveness of the invention. Example 2

[0056] This embodiment, based on the system described in Embodiment 1, provides a control method for an autonomous mobile charging robot. This method includes six main steps: mapping and localization, global path planning, visual perception, trajectory prediction, local path planning, and docking and charging. The following detailed description is provided in conjunction with a specific experimental scenario.

[0057] I. Experimental Preparation Select an area of ​​30m 2An indoor parking lot was used as the test scenario, which included three parking spaces, several fixed obstacles, and randomly appearing pedestrians as dynamic obstacles. The mobile charging robot was assembled according to the configuration of Example 1, and the charging interface was marked with a red circular symbol with a diameter of 10cm.

[0058] The testing equipment includes a high-precision positioning device (error ±0.1cm), a lux meter, a stopwatch, and a mobile phone (Android system, with a test APP installed).

[0059] II. Specific Implementation Process of the Method Step 1: Mapping and Positioning After the robot starts, the host computer runs the gmapping function package. The LiDAR scans the surrounding environment at a frequency of 5-8Hz, and the IMU provides robot posture data. The system constructs a two-dimensional grid map based on a filtering algorithm, with a map resolution of 0.05 meters per pixel. During navigation, the AMCL algorithm matches the real-time LiDAR scan data with the pre-built map to estimate the robot's current position and orientation angle, keeping the positioning error within 2cm.

[0060] Step 2: Global Path Planning Users select a target parking space via a mobile app and click "One-Click Start." The coordinates of the target parking space are sent to a lower-level device via Bluetooth, which then forwards them to a higher-level device. Upon receiving the charging command, the higher-level device invokes the global path planner. The global path planner uses the A* algorithm, starting from the robot's current position and ending at the target parking space, to calculate the optimal collision-free path on the map using heuristic functions (such as Manhattan distance or Euclidean distance). This generates a series of path point sequences, which the robot chassis then controls to move along.

[0061] Step 3: Visual Perception The perception unit continues to operate during the robot's movement and after reaching the target area. A monocular camera captures images, which are resized to 320×320 pixels. A lightweight deep learning object detection network performs inference on each frame. Specifically, the object detection network uses YOLOv26 and incorporates the following features: (a) End-to-end NMS-free inference architecture: YOLOv26 employs an innovative dual-head training mechanism, using both a "one-to-many" head (providing dense supervision) and a "one-to-one" head (learning the final prediction) during the training phase, while retaining only the "one-to-one" head during the inference phase. This completely eliminates the non-maximum suppression (NMS) post-processing step, achieving true end-to-end inference. It decouples model inference latency from scene complexity, maintaining constant low latency even in dense target scenes, thus addressing the pain point of uncertain inference time in traditional detectors in areas with dense dynamic obstacles (such as pedestrian-heavy areas in parking lots).

[0062] (ii) Small Target Perceptual Label Assignment (STAL): For small targets like charging ports (approximately 10cm in diameter, typically less than 8 pixels in an image), YOLOv26 introduces the STAL mechanism, which forces the allocation of at least 4 anchor points for supervised training of small targets. This ensures that the model can stably learn the feature representation of small targets, significantly improving the detection recall rate of long-distance charging ports.

[0063] (iii) ProgLoss: During training, the loss weights of "one-to-many" and "one-to-one" heads are dynamically adjusted. In the early stage, dense supervision is emphasized to accelerate convergence, and in the later stage, end-to-end prediction-aligned inference behavior is emphasized, making the model more stable and converges faster when trained on the embedded platform (Raspberry Pi 4B).

[0064] (iv) Multiscale Linear Attention Enhancement (MSLA): An improved MSLA module is introduced into the YOLOv26 backbone network, which utilizes parallel 4-branch depthwise convolutions ( , , , Multi-scale features are extracted and combined with a linear attention mechanism to capture global long-range dependencies, thereby enhancing the ability to represent features of the charging interface at different distances.

[0065] (v) Multi-channel refiner (MDCR): The MDCR module is integrated into the YOLOv26 neck network, dividing the features along the channel dimension into four heads, each with a different expansion rate. The depthwise separable convolution, through channel interleaving and point convolution fusion, achieves refined extraction of multi-receptive field features, further improving the localization accuracy of small targets.

[0066] Furthermore, the lightweight deep learning object detection network is also equipped with a multi-scene adaptive visual perception module. This module connects a lightweight scene classification branch in parallel before the backbone network. This branch consists of three layers of depthwise separable convolutions and global average pooling, outputting scene type labels and lighting environment quality scores in real time. Based on the scene classification results, the dynamic feature enhancement module automatically selects and activates the corresponding processing branches: for low-light scenes, the Retinex-Net branch is activated to decouple the illumination and reflection components and enhance the reflection component; for strong backlight scenes, local histogram equalization and gamma correction are activated to dynamically compress overexposed areas and restore shadow details; for dynamically blurred scenes, deformable convolution branches are activated to adaptively adjust the sampling point positions to compensate for motion blur; for small target distant scenes, a high-resolution upsampling branch is added to the feature pyramid network to enhance the feature response of small targets; for normal lighting scenes, no enhancement branches are activated to maintain the highest inference speed.

[0067] Finally, the perception unit multiplies the ambient light quality score by the class confidence score output by the detection head, and then performs temporal smoothing of the center coordinates of the detection boxes in consecutive frames using Kalman filtering to obtain the visual perception reliability score. This score reflects the reliability of the current visual perception, ranging from 0 to 1. The perception unit also outputs dynamic obstacle detection results, including obstacle category and location information.

[0068] Step 4: Trajectory Prediction When the visual perception module detects a dynamic obstacle, such as a pedestrian, the system uses a Kalman filter to estimate the motion state of each obstacle. Based on the detection results of multiple consecutive frames, the Kalman filter infers the obstacle's position, velocity, and acceleration, and extrapolates a sequence of predicted trajectory points for the next second. Each predicted point includes coordinate information and a corresponding timestamp.

[0069] Step 5: Local Path Planning The improved TEB algorithm in the planning unit performs local path planning at a frequency of 20 Hz. The TEB algorithm obtains smooth and collision-free local paths by optimizing the cost function.

[0070] 5.1 Basic Cost Function Structure The cost function J of the TEB algorithm consists of the following five terms: path length term. To make the trajectory as short as possible; static obstacle distance term The dynamic obstacle distance term helps the robot avoid fixed obstacles such as walls and pillars; the dynamic obstacle distance term helps the robot avoid moving obstacles such as pedestrians and vehicles; and the acceleration change term helps the robot avoid moving obstacles such as pedestrians and vehicles. Constraints on the smoothness of motion; velocity variation term To constrain the continuity of motion, each term is multiplied by its corresponding weighting coefficient and then summed. The optimal local trajectory is obtained by minimizing this sum.

[0071] 5.2 Spatiotemporal Correlation Penalty Term Based on Predicted Trajectory After the Kalman filter in step four generates the sequence of predicted trajectory points for the next second, the planning unit incorporates these predicted points into the dynamic obstacle distance term. The dynamic obstacle distance term is no longer simply based on the current position of the obstacle, but instead employs a spatiotemporal correlation penalty term. This penalty term considers both the spatial distance and temporal proximity between the robot's pose and the predicted trajectory point, and its expression is: in, For spatiotemporal correlation penalty terms, The preset dynamic obstacle weights, For safe distance threshold, The robot pose output by the TEB planner. For the predicted dynamic obstacle trajectory points, and Robot pose and trajectory points The corresponding timestamp, The attenuation coefficient is... This reflects the characteristic that forecast uncertainty increases over time.

[0072] The physical meaning of this penalty term is: when the spatial distance between the robot's pose and a predicted trajectory point is less than a safe distance threshold, a positive cost is incurred; and the closer to the current moment in time, the larger the exponential decay term and the more significant the cost, reflecting the characteristic that prediction uncertainty increases with time. By summing i and j twice, it is ensured that each pose point of the robot is evaluated for spatiotemporal conflict with each predicted trajectory point.

[0073] In specific implementation, It is directly used as the dynamic obstacle distance term in the TEB cost function, replacing the simple distance penalty term in the traditional TEB that is only based on the current position of the obstacle.

[0074] 5.3 Adaptive Adjustment of Weighting Coefficients In this invention, the weight coefficients are not fixed, but are modulated step by step through the following three levels, enabling the robot to automatically adjust its behavior strategy according to changes in the environment.

[0075] First layer: Adaptive allocation of sensor trust levels. The system scores reliability based on visual perception. Based on the confidence interval, the confidence allocation ratio of the visual sensor, LiDAR, and IMU is adjusted in stages. When When the light level exceeds a first preset threshold, such as 0.7, it is considered that the lighting is good and the scene is clear. The visual sensor obtains the dominant trust level, and its trust level is set to a relatively high value, such as 0.8. The LiDAR trust level is set to the highest value, such as 0.9, and the IMU trust level is set to a medium value, such as 0.7. When the value falls below a second preset threshold, such as 0.4, it is considered to be in a difficult scenario such as low light or dynamic blur. The system then lowers the confidence level of the visual sensor to a lower value, such as 0.3, while raising the confidence level of the LiDAR to the highest value, such as 0.95, and the confidence level of the IMU to a higher value, such as 0.85. This adjustment layer ensures the reliability of the perception information relied upon by the planner.

[0076] The second layer: Weighted adjustment via a switch. When the visual perception module detects a dynamic obstacle, the system adjusts the weighting coefficient of that obstacle. The value was increased from the baseline of 0.20 to a higher value in the range of 0.35 to 0.45, while the weight of the path length term was also increased. The weights are reduced from the baseline value of 0.35 to a lower value within the range of 0.25 to 0.30 to implement a "safety-first" obstacle avoidance strategy. When an open area without dynamic obstacles is detected, the weights are restored to the baseline value. When the robot is in a narrow passage, the time-optimal weight of the velocity change term is increased to suppress drastic changes in angular velocity and prevent collision risks caused by sharp turns.

[0077] The third layer: Continuous nonlinear adjustment based on scene complexity. Building upon the first two layers, the system further refines the dynamic obstacle avoidance intensity and time optimality preference through continuous fine-tuning. The system first defines the scene complexity index C: in, The number of dynamic targets detected. To calculate the spatial clustering degree of the target using kernel density estimation, Assess the reliability of visual perception. , , The preset weighting coefficients are 0.4, 0.3, and 0.3, respectively, and are determined using the analytic hierarchy process.

[0078] Then, the system assigns weights to dynamic obstacles respectively. And time optimality weight Nonlinear adjustment is performed according to the following formula: in, This serves as the baseline value for the dynamic obstacle weight. This serves as the baseline value for the time-optimal weight. To adjust the amplitude coefficient, To reduce the coefficient, This is the saturation threshold.

[0079] In this embodiment, Take 0.20, Set the value to 0.25, adjust the amplitude coefficient k to 0.8, reduce the coefficient η to 0.6, and set the saturation threshold. The value is 0.8. When the value of C is small, the tanh term is close to 0. , When the value of C increases, Gradually rising, Gradually decreasing.

[0080] 5.4 Handling Extreme Situations When visual perception reliability score When the visual field deviates below a preset threshold, such as 0.2 (in extreme darkness, lens obstruction, or severe motion blur), the system determines that vision is completely unusable and triggers blind docking mode. In this mode, the planning unit completely abandons visual guidance and no longer relies on visual perception reliability scores for weight adjustment. Instead, it directly completes docking positioning based on the robotic arm posture recorded in historical successful docking records and the fixed features of the parking space identified by LiDAR (such as the outlines of vehicles on both sides and the position of blocks). It performs a blind insertion according to the memorized posture and makes fine adjustments within a range of ±5° through feedback from the photosensitive sensor to ensure that the charging task can still be completed in extreme scenarios.

[0081] 5.5 Complete Cost Function Expression Substituting the above spatiotemporal correlation penalty term and each adaptive weight, the TEB cost function is finally expressed as: in, , , These are the weighting coefficients for the path length term, the static obstacle distance term, and the acceleration change term, respectively. , These are the adjusted dynamic obstacle weights and time-optimal weights, respectively. As a preset positive number, For path length, The distance from the robot to the static obstacle. For acceleration changes, This represents the change in speed.

[0082] The TEB optimizer uses the aforementioned cost function J as the evaluation metric to solve for the optimal pose sequence and velocity command through online optimization. The overall system replanning response time does not exceed 1 second. The planning unit sends the optimized linear velocity and angular velocity commands to the lower-level machine via ROS topics. The lower-level machine drives the motor to operate through PID control to achieve obstacle avoidance navigation or docking actions.

[0083] Step Six: Connect for Charging Once the robot reaches the vicinity of the target parking space through the aforementioned navigation and obstacle avoidance processes, the sensing unit focuses on the precise identification of the charging interface. The robotic arm's chassis servo motors reciprocate within a ±30° range until a red circular marker is detected. The system calculates the pixel deviation between the center of the detection frame and the preset center point of the image, converts it into the spatial pose deviation of the robotic arm's end effector, and drives the servo motors to perform closed-loop adjustments until the spatial pose error does not exceed 2mm. At this point, the simulated charging gun is inserted into the interface. After the photosensitive sensor detects the insertion signal, the charging process begins. Once fully charged, the system sends a "charging complete" signal via Bluetooth, and the robot automatically returns to the standby point.

[0084] III. Experimental Results Using the above method, in mixed tests encompassing four scenarios—normal lighting, low lighting, strong backlighting, and dynamic blurring—the overall charging success rate of this invention reaches 96.5%, the dynamic obstacle avoidance response time is reduced to less than 0.4 seconds, and the path deviation after obstacle avoidance is less than 1.2 centimeters. Compared to traditional fixed-weight schemes, this significantly improves the reliability and intelligence level of the mobile charging robot in complex environments.

[0085] Finally, it should be noted that any parts of this invention not described in detail are prior art. Those skilled in the art will understand that the above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A control method for an autonomous mobile charging robot, characterized in that, Includes the following steps: Mapping and localization: Constructing an environmental map based on multimodal sensor data and locating the robot's pose in real time; Global path planning: After receiving the charging command, global path planning is performed based on the pre-built environment map to control the robot to move to the target area; Visual perception: Utilizing a lightweight deep learning object detection network deployed on edge computing units, the acquired images are processed in real time to detect charging ports and dynamic obstacles, and output visual perception reliability scores and dynamic obstacle detection results. Trajectory prediction: Predict the trajectory of detected dynamic obstacles and generate predicted trajectory information; Local path planning: Based on the robot's current pose, global path, LiDAR data, the dynamic obstacle detection results, and predicted trajectory information, local path planning is performed using an improved TEB algorithm; wherein, the cost function of the improved TEB algorithm includes a spatiotemporal correlation penalty term generated based on the predicted trajectory information, and the planning parameters are dynamically adjusted according to the output results of the visual perception step, and the planning parameters include at least the weight coefficient of the dynamic obstacle distance term. Docking and charging: When the robot arrives at the target area, it drives the robot's actuator to complete docking and charging based on the charging interface detection results; The spatiotemporal correlation penalty term is: in, For spatiotemporal correlation penalty terms, The preset dynamic obstacle weights, For safe distance threshold, The robot pose output by the TEB planner. For the predicted dynamic obstacle trajectory points, and Robot pose and trajectory points The corresponding timestamp, The attenuation coefficient is... This reflects the characteristic that forecast uncertainty increases over time; The cost function of the improved TEB algorithm includes a path length term, a static obstacle distance term, a dynamic obstacle distance term, an acceleration change term, and a velocity change term. The dynamic obstacle distance term employs a spatiotemporal correlation penalty term based on the predicted trajectory. The optimization objective of the cost function is to minimize the following expression: in, , , These are the weighting coefficients for the path length term, the static obstacle distance term, and the acceleration change term, respectively. , These are the adjusted dynamic obstacle weights and time-optimal weights, respectively. As a preset positive number, For path length, The distance from the robot to the static obstacle. For acceleration changes, This represents the change in speed.

2. The control method for an autonomous mobile charging robot according to claim 1, characterized in that, The visual perception reliability score is calculated as follows: the input image is classified into scenes, the scene type and light environment quality score are output, the light environment quality score is multiplied by the category confidence score output by the target detection network, and the comprehensive reliability score is obtained after time-series smoothing filtering.

3. The control method for an autonomous mobile charging robot according to claim 1, characterized in that, The trajectory prediction of the dynamic obstacle uses a Kalman filter to estimate the motion state of the dynamic obstacle based on the detection results of multiple consecutive frames, and extrapolates the predicted trajectory point sequence in the future time period.

4. The control method for an autonomous mobile charging robot according to claim 3, characterized in that, Dynamically adjusting planning parameters based on the output of the visual perception step includes at least one of the following methods: (a) Sensor Trust Level Adaptive Allocation: Based on the confidence interval of the visual perception reliability score, the trust level allocation ratio of the visual sensor, LiDAR and inertial measurement unit is adjusted in stages; wherein, when the visual perception reliability score is in the high confidence interval, the visual sensor obtains the dominant trust level; when the visual perception reliability score is in the low confidence interval, the trust level of the visual sensor is suppressed and the LiDAR obtains the dominant trust level. (b) Planning weight switch adjustment: When a dynamic obstacle is detected, increase the weight coefficient of the dynamic obstacle distance term and correspondingly decrease the weight coefficient of the path length term; when in a narrow passage, increase the weight coefficient of the speed change term; (c) Blind docking mode: When the visual perception reliability score is lower than the preset threshold, visual guidance is completely abandoned, and docking is completed based on historical docking records and lidar positioning.

5. The control method for an autonomous mobile charging robot according to claim 4, characterized in that, Dynamic adjustment of planning parameters also includes continuous nonlinear adjustment based on scenario complexity: Scene complexity metrics Defined as: in, The number of dynamic targets detected. To estimate the spatial clustering degree of the target using kernel density estimation, Assess the reliability of visual perception. , , These are preset weighting coefficients; Adjust dynamic obstacle weights using the following formula And time optimality weight : in, This serves as the baseline value for the dynamic obstacle weight. This serves as the baseline value for the time-optimal weight. To adjust the amplitude coefficient, To reduce the coefficient, This is the saturation threshold.

6. The control method for an autonomous mobile charging robot according to claim 1, characterized in that, The lightweight deep learning object detection network has at least one of the following characteristics: End-to-end non-maximum suppression inference architecture: The network adopts a dual-head training mechanism, using both one-to-many and one-to-one supervision during the training phase, and retaining only the output of the one-to-one head during the inference phase, without the need for non-maximum suppression post-processing. Small target perception label allocation mechanism: Multiple anchor points are assigned to small targets at the charging interface for supervised training; Multi-scale linear attention enhancement module: Introduces multi-branch deep convolution in the backbone network to extract multi-scale features in parallel, and combines linear attention to capture global long-range dependencies; Multi-diffractory channel refiner: The neck network is divided into multiple heads along the channel dimension, and depthwise separable convolutions with different dilation rates are applied to each head. Through channel interleaving and point convolution fusion, refined extraction of multi-receptive field features is achieved. Multi-scene adaptive visual perception module: includes parallel scene classification branch and dynamic feature enhancement branch. Based on the scene classification results, it adaptively selects low light enhancement, backlight correction, dynamic blur compensation or high-resolution upsampling processing for small targets, and outputs a visual perception reliability score.

7. The control method for an autonomous mobile charging robot according to claim 1, characterized in that, The lightweight deep learning object detection network, after channel pruning and INT8 quantization, is deployed on a Raspberry Pi edge computing platform and communicates data with zero copy via shared memory.

8. A control system for an autonomous mobile charging robot, used to implement the control method for an autonomous mobile charging robot according to any one of claims 1-7, characterized in that, include: The main control unit adopts a hierarchical architecture of upper computer and lower computer. The upper computer is used to run SLAM mapping, path planning and visual recognition algorithms, while the lower computer is used for low-level drive control. The multimodal sensor unit, including lidar, inertial measurement unit and monocular camera, is used to collect environmental perception data and robot status data; The drive execution unit, including a mobile chassis, a motor drive module, and a robotic arm, is used to drive the robot to move and complete the charging interface docking according to the instructions of the main control unit. The interaction unit includes a voice module, a display module, and a remote control module, which are used to receive user commands and provide feedback on the robot's status. The host computer is equipped with a lightweight deep learning object detection network that has undergone pruned quantization and runs an improved TEB algorithm.

Images