A model training method and a robot obstacle avoidance method

Abstract

The application belongs to the technical field of robots, and particularly relates to a model training method and a robot obstacle avoidance method, comprising the following steps: acquiring an expert demonstration data set; constructing a training sample from the data set, wherein the training sample comprises a historical point cloud sequence, a historical ontology state sequence and an expert action sequence; preprocessing the historical point cloud sequence, the expert action sequence and the historical ontology state sequence; adding noise to the expert action sequence to generate a noisy action; training a noise prediction network based on the historical point cloud sequence, the historical ontology state sequence and the noisy action to predict the added noise; and optimizing parameters of the noise prediction network according to the difference between the predicted noise and the added noise. The application can enhance the adaptability of a cleaning robot in a curtain wall cleaning environment, and improve the cleaning efficiency and safety performance of the robot.

Description

Technical Field

[0001] This invention belongs to the technical field of robotics, specifically relating to a model training method and a robot obstacle avoidance method. Background Technology

[0002] Current curtain wall cleaning robots mainly rely on pre-programmed methods and sensors to achieve obstacle avoidance. These solutions can usually only handle simple obstacles, but they are difficult to effectively avoid in complex environments, such as protruding window frames, glass seams, broken glass segments, and glass holes.

[0003] In recent years, some studies have begun to use machine learning and visual recognition technologies to assist obstacle avoidance through binocular cameras and deep learning algorithms. These technical solutions typically employ convolutional neural networks (CNNs) for image classification and object detection, but lack an understanding of 3D environmental information, resulting in poor obstacle avoidance performance in complex scenes. Summary of the Invention

[0004] This application discloses a model training method and a robot obstacle avoidance method, which can improve the obstacle avoidance ability of cleaning robots.

[0005] The first aspect of this application discloses a model training method, including the following steps: S1. Obtain the expert demonstration dataset; S2. Construct training samples from the dataset, the training samples including historical point cloud sequences, historical ontology state sequences, and expert action sequences; S3. Preprocess the historical point cloud sequence, the expert action sequence, and the historical ontology state sequence; S4. Add noise to the expert action sequence to generate noisy actions; S5. Based on the historical point cloud sequence, the historical ontology state sequence, and the noisy action, train a noise prediction network to predict the added noise; S6. Optimize the parameters of the noise prediction network based on the difference between the predicted noise and the added noise.

[0006] As an optional implementation, in the first aspect of the embodiments of this application, the preprocessing of the historical point cloud sequence, the expert action sequence, and the historical ontology state sequence includes the following steps: The historical point cloud sequence is voxelized and uniformly sampled; The expert action sequence and the historical ontology state sequence are normalized.

[0007] As an optional implementation, in the first aspect of the embodiments of this application, training a noise prediction network based on the historical point cloud sequence, the historical ontology state sequence, and the noisy action to predict the added noise includes the following steps: The historical point cloud sequence is input into the first encoder to extract visual feature vectors; The historical ontology state sequence is input into the second encoder to extract motion feature vectors; The noisy action, the intensity of the added noise, the visual feature vector, and the motion feature vector are input into a UNet network to predict the noise added to the expert action sequence.

[0008] As an optional implementation, in the first aspect of the embodiments of this application, optimizing the parameters of the noise prediction network based on the difference between the predicted noise and the added noise includes the following steps: Calculate the mean square error between the predicted noise output by the UNet network and the added noise; Based on the mean square error, the weights of the first encoder, the second encoder, and the UNet network are updated using backpropagation and gradient descent algorithms.

[0009] The second aspect of this application discloses a robot obstacle avoidance method, which employs a model training method disclosed in the first aspect of this application, and includes the following steps: Data acquisition involves obtaining data from tasks performed by robots operated by human experts and generating expert demonstration datasets. Model training: Based on the expert demonstration dataset, obstacle avoidance strategies are determined. Deployment and reasoning: Obtain the current point cloud sequence and current body state sequence of the robot's environment, input the obstacle avoidance strategy, and generate a future action sequence.

[0010] As an optional implementation, in the second aspect of the embodiments of this application, the acquisition of data on tasks completed by human experts operating robots includes the following steps: Expert remote operation allows human operators to remotely control the trainer aircraft from a virtual simulation environment, using the aircraft's perception perspective, and capture motion data of each joint during mission execution.

[0011] As an optional implementation, in the second aspect of the embodiments of this application, the generation of the expert demonstration dataset includes the following steps: Motion mapping involves mapping the captured joint motion data of the trainer machine onto the robot using kinematic algorithms. Data is collected synchronously to control the robot to perform the mapped actions, and sensor data and body status data of the robot are collected synchronously. Dataset construction involves processing the synchronous data collected from multiple tasks to generate an expert demonstration dataset; wherein, the expert demonstration dataset includes historical point cloud sequences, historical ontology state sequences, and expert action sequences.

[0012] As an optional implementation, in a second aspect of the embodiments of this application, determining the obstacle avoidance strategy based on the expert demonstration dataset includes the following steps: Training samples are constructed from the expert demonstration dataset, and the training samples include historical point cloud sequences, historical ontology state sequences, and expert action sequences; The historical point cloud sequence is voxelized and uniformly sampled; Normalize the expert action sequence and the historical ontology state sequence; Noise is added to the expert action sequence to generate noisy actions; The historical point cloud sequence is input into the first encoder to extract visual feature vectors; The historical ontology state sequence is input into the second encoder to extract motion feature vectors; The noisy action, the intensity of the added noise, the visual feature vector, and the motion feature vector are input into a UNet network to predict the added noise. The mean square error between the predicted noise output by the UNet network and the added noise is calculated, and the weights of the first encoder, the second encoder, and the UNet network are updated based on this error to obtain the obstacle avoidance strategy.

[0013] As an optional implementation, in the second aspect of the embodiments of this application, the step of obtaining the current point cloud sequence and the current body state sequence of the robot's environment, inputting the obstacle avoidance strategy, and generating a future action sequence includes the following steps: To obtain current observations, the robot acquires the real-time point cloud of the current environment and reads the current joint status; Iterative denoising and action generation: Based on current observations, future action sequences are iteratively generated using the trained obstacle avoidance strategy.

[0014] As an optional implementation, in a second aspect of the embodiments of this application, after generating the future action sequence, the method further includes: Action execution and closed-loop control: Obtain the action of the first time step of the future action sequence and send it to the underlying controller for execution; After the robot completes a step, it reads new sensor data and repeats the steps to generate the future action sequence based on the new sensor data, forming a closed-loop control system.

[0015] The third aspect of this application discloses an obstacle avoidance cleaning method for a curtain wall cleaning robot, which employs a model training method disclosed in the first aspect of this application or a robot obstacle avoidance method disclosed in the second aspect, and includes the following steps: When the robot detects a crossable obstacle in front of it, it is controlled to execute a preset obstacle-crossing action procedure to cross the obstacle according to the obstacle avoidance strategy. In a non-obstacle-crossing state, the robot is controlled to perform an automatic cleaning process, which includes lowering the cleaning mechanism and driving it to perform cleaning movements along the curtain wall surface. When it is necessary to change the cleaning direction, control the robot to perform a turning process.

[0016] The fourth aspect of this application discloses a cleaning robot, including a memory storing executable program code and a processor coupled to the memory; wherein the processor calls the executable program code stored in the memory to execute the method disclosed in any of the second aspects of this application.

[0017] The fifth aspect of this application discloses a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method disclosed in any of the first or second aspects of this application.

[0018] Compared with related technologies, the embodiments of this application have the following beneficial effects: This invention includes the following steps: acquiring an expert demonstration dataset; constructing training samples from the dataset, the training samples including historical point cloud sequences, historical ontology state sequences, and expert action sequences; voxelizing and uniformly sampling the historical point cloud sequences; normalizing the expert action sequences and the historical ontology state sequences; adding noise to the expert action sequences to generate noisy actions; training a noise prediction network based on the historical point cloud sequences, the historical ontology state sequences, and the noisy actions to predict the added noise; and optimizing the parameters of the noise prediction network based on the difference between the predicted noise and the added noise. This invention can be used to train obstacle avoidance models for cleaning robots, improving the obstacle avoidance capabilities of cleaning robots, enhancing their adaptability in curtain wall cleaning environments, and improving their cleaning efficiency and safety performance. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is one of the flowcharts illustrating a model training method disclosed in an embodiment of this application; Figure 2 This is a second schematic flowchart of a model training method disclosed in an embodiment of this application; Figure 3 This is the third flowchart of a model training method disclosed in an embodiment of this application; Figure 4 This is the fourth flowchart of a model training method disclosed in the embodiments of this application; Figure 5 This is a flowchart illustrating a robot obstacle avoidance method disclosed in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of a curtain wall cleaning robot disclosed in an embodiment of this application; Figure 7 This is a schematic flowchart of an obstacle avoidance cleaning method for a curtain wall cleaning robot disclosed in an embodiment of this application; Figure 8 This is a flowchart illustrating another robot obstacle avoidance method disclosed in an embodiment of this application; Figure 9 This is a schematic diagram of the structure of a cleaning robot disclosed in an embodiment of this application.

[0021] The components are as follows: 1. First adsorption mechanism; 100. First suction cup; 101. First suction cup lifting motor; 2. Second adsorption mechanism; 21. Front suction cup; 22. Front suction cup lifting motor; 221. First guide post; 23. Rear suction cup; 24. Rear suction cup lifting motor; 241. Second guide post; 3. Pull rope motor; 4. Forward and backward moving motor; 5. Central rotary motor; 6. Cleaning U-axis rotary motor; 61. Limiting component; 7. Cleaning W-axis rotary motor; 8. Cleaning mechanism; 9. Frame; 10. Guide component; 11. Connecting component; 12. Memory; 13. Processor. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] The terms “comprising” and “having”, and any variations thereof, in this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0024] In the embodiments of this application, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0025] The technical solution of this application will be described in detail below with reference to specific embodiments.

[0026] Please see Figure 1 A model training method includes the following steps: S1. Obtain the expert demonstration dataset. The expert demonstration dataset is obtained by recording high-quality data of human experts operating robots to complete tasks, which serves as the data basis for training the model. S2. Construct training samples from the dataset. The training samples include historical point cloud sequences, historical ontology state sequences, and expert action sequences. S3. Preprocess the historical point cloud sequence, expert action sequence, and historical ontology state sequence; S4. Add random noise of preset intensity to the expert action sequence to generate noisy actions. The added noise can be Gaussian noise. S5. Based on historical point cloud sequences, historical ontology state sequences, and noisy actions, train a noise prediction network to predict added noise. S6. Optimize the parameters of the noise prediction network based on the difference between the predicted noise and the added noise; S7. Determine the obstacle avoidance strategy.

[0027] In some embodiments, the historical point cloud sequence, expert action sequence, and historical ontology state sequence are preprocessed, including the following steps: S31. Perform voxelization and uniform sampling on the historical point cloud sequence; S32. Normalize the expert action sequence and the historical ontology state sequence.

[0028] In some embodiments, a noise prediction network is trained based on historical point cloud sequences, historical ontology state sequences, and noisy actions to predict added noise, including the following steps: S51. Input the historical point cloud sequence into the first encoder to extract the visual feature vector. The first encoder can be a PointNet++ encoder or a Pyramid ConvNet encoder. S52. Input the historical ontology state sequence into the second encoder to extract motion feature vectors, wherein the second encoder can be an MLP encoder; S53. Input the noisy action, the intensity of the added noise, the visual feature vector, and the motion feature vector into a UNet network to predict the noise added to the expert action sequence.

[0029] In some embodiments, optimizing the parameters of the noise prediction network based on the difference between the predicted noise and the added noise includes the following steps: S61. Calculate the mean square error between the predicted noise and the added noise in the output of the UNet network; S62. Based on the mean square error, update the weights of the first encoder, the second encoder, and the UNet network through backpropagation and gradient descent algorithms.

[0030] This application discloses a robot obstacle avoidance method, which employs a model training method disclosed in the above embodiments, and includes the following steps: Step 100: Data Acquisition. Obtain data on tasks completed by human experts operating the robot and generate an expert demonstration dataset. Step 200: Model training. Based on the expert demonstration dataset, determine the obstacle avoidance strategy. Step 300: Deployment and Reasoning. Obtain the current point cloud sequence and current body state sequence of the robot's environment, input the obstacle avoidance strategy, and generate the future action sequence.

[0031] In some embodiments, data on tasks performed by robots operated by human experts are acquired to generate an expert demonstration dataset, including the following steps: Step 101: Expert remote operation. In a virtual simulation environment, a human operator remotely controls the trainer machine through the trainer machine's perception perspective and captures the motion data of each joint of the trainer machine during the execution of the task. The perception perspective of the trainer machine can be the perspective of the trainer machine's binocular camera, and the trainer machine is a training robot. Step 102, Motion Mapping: The captured joint motion data of the trainer machine is mapped onto the robot using a kinematic algorithm; Step 103: Data synchronous acquisition, control the robot to execute the mapped actions, and synchronously acquire the robot's sensor data and body status data; Step 104: Dataset Construction. The synchronous data collected from multiple tasks is processed to generate an expert demonstration dataset. The samples in the expert demonstration dataset include historical point cloud sequences, historical body state sequences, and expert action sequences. Specifically, historical point cloud sequences and historical body state sequences are generated based on the synchronously collected sensor data and body state data of the robot. Expert action sequences are generated based on the motion data of each joint of the trainer machine during the execution of the task.

[0032] In some embodiments, determining an obstacle avoidance strategy based on an expert demonstration dataset includes the following steps: Step 201: Construct training samples from the expert demonstration dataset. The training samples include historical point cloud sequences, historical ontology state sequences, and expert action sequences. Step 202: Voxelize and uniformly sample the historical point cloud sequence; Step 203: Normalize the expert action sequence and the historical ontology state sequence; Step 204: Add random noise of preset intensity to the expert action sequence to generate noisy actions; Step 205: Input the historical point cloud sequence into the first encoder to extract the visual feature vector; Step 206: Input the historical ontology state sequence into the second encoder to extract the motion feature vector; Step 207: Input the noisy action, the intensity of the added noise, the visual feature vector, and the motion feature vector into a UNet network to predict the added noise; Step 208: Calculate the mean square error between the predicted noise and the added noise in the output of the UNet network, and update the weights of the first encoder, the second encoder and the UNet network based on this mean square error. Repeatedly execute steps 204 to 208 above. As training progresses, the model's ability to predict noise continuously improves. When the model training is completed (converged), the final obstacle avoidance strategy is obtained.

[0033] In some embodiments, the expert demonstration dataset is split and extracted to obtain a training dataset and a test dataset. The training dataset is used to train the obstacle avoidance strategy, and the test dataset is used to verify the obstacle avoidance strategy.

[0034] In some embodiments, the current point cloud sequence and current body state sequence of the robot's environment are obtained, the obstacle avoidance strategy is input, and a future action sequence is generated, including the following steps: Step 301: Obtain current observation. The robot obtains the real-time point cloud of the current environment through a binocular depth camera and reads the current joint state of the robot. Step 302: Iteratively denoise and generate actions. Based on the current observations, iteratively generate future action sequences using the trained obstacle avoidance strategy.

[0035] In some embodiments, after generating the sequence of future actions, the method further includes: Step 303, Action Execution and Closed-Loop Control: Obtain the action of the first time step of the future action sequence and send it to the underlying motor controller for execution; Step 304: After the robot completes one step, it reads new sensor data and repeats the deployment and reasoning steps based on the new sensor data to form a closed-loop control system.

[0036] The obstacle avoidance method of this invention can be applied to autonomous obstacle avoidance decision-making for curtain wall robots, and includes three stages: Phase 1, Data Collection: This phase involves generating an expert demonstration dataset by recording high-quality data of human experts operating the coach robot to complete tasks. This includes the following steps: Teleoperation: This involves a human operator manipulating a coaching robot in a virtual environment using a binocular camera. The operator controls the robot with their own movements, precisely capturing the motions of each joint.

[0037] Kinematic Mapping: The captured joint movements of the coach robot are mapped onto the trainee robot using a kinematic algorithm (Relaxed IK). For example, the landing position of the coach robot when crossing obstacles is converted into the target position of multiple suction cups and joints of the robot, and then the joint angles of the robot are solved.

[0038] Data Synchronization Recording: The system synchronously records all sensor data of the robot when performing the above actions, forming an observation-action pair. Observation includes vision (raw 3D point cloud from binocular depth camera) and proprioception (internal states such as joint angles and motor torque), while action is the target joint position sequence generated by the trainer machine.

[0039] Dataset construction: Store the data trajectories (containing multiple observation-action pairs) of multiple successful task completions to form an expert demonstration dataset. The expert demonstration dataset includes historical point cloud sequences (visual), historical ontology state sequences (proprioception), and expert action sequences (actions).

[0040] Phase Two, Model Training, involves teaching the model to mimic the behavior of experts. This includes the following steps: Data loading and preprocessing: Short time series data (such as 8-step historical observations and 16-step future actions) are randomly sampled from the expert demonstration dataset. The historical point cloud sequence is voxelized and uniformly sampled. The uniform sampling is fixed at N points, where N=4096. The expert action sequence and historical ontology state data are normalized.

[0041] Diffusion Process: Forward noise addition, i.e., randomly selecting Gaussian noise with noise intensity t and adding Gaussian noise to the historical expert action sequence a_0 according to a predetermined ratio to generate noisy action a_t.

[0042] Conditional encoding: The historical point cloud sequence is encoded by a Pyramid ConvNet encoder to extract compressed visual feature vectors, while the historical ontology state sequence is encoded by an MLP encoder.

[0043] Noise prediction: The noisy action a_t, the noise intensity t, the encoded historical point cloud sequence, and the encoded historical ontology state sequence are input into a UNet network. The core task of the UNet network is to predict the noise ε_θ added to the expert action sequence.

[0044] Loss calculation and optimization: The mean squared error (MSE Loss) between the predicted noise ε_θ and the true noise ε is calculated. The weights of the UNet network and encoder are updated through backpropagation and gradient descent (using the AdamW optimizer), so that the noise prediction network composed of the UNet network and encoder can more accurately predict the added noise, thereby improving its denoising performance.

[0045] Phase 3, Deployment & Inference: The trained model is run on a real robot. This includes the following steps: Acquire current observations: The robot acquires real-time point clouds of the current environment through a binocular depth camera and simultaneously reads the current joint status.

[0046] Iterative denoising and action generation: Starting with a set of random Gaussian noise action sequences, combined with the conditional features encoded by historical observations, the input is a trained diffusion UNet network; the diffusion UNet network predicts the current noise and performs denoising operations, updating the action sequence (resulting in a cleaner action sequence); the above process is repeated T times (e.g., 10 times), gradually denoising and optimizing the action sequence, and finally outputting a reasonable future action sequence (gradually refining a reasonable action sequence).

[0047] Action execution and closed-loop control: Take the first time step action in the generated future action sequence and send it to the underlying motor controller for execution; after execution, reacquire the latest sensor data and repeat the above steps from acquiring the current observation to action execution and closed-loop control.

[0048] After the robot completes a step, it reads new sensor data and repeats the entire process, forming a closed-loop control system that enables real-time self-adjustment to environmental changes.

[0049] Using the methods described above, robots can accurately understand the shapes of obstacles in three-dimensional space, autonomously decide whether to avoid or overcome obstacles, and improve cleaning efficiency and safety in complex environments.

[0050] In some embodiments, a method for obstacle avoidance cleaning with a curtain wall cleaning robot includes the following steps: Step S100: When the robot detects a crossable obstacle in front of it, it controls the robot to execute a preset obstacle-crossing action procedure to cross the obstacle according to the obstacle avoidance strategy. The process of controlling the robot to execute a preset obstacle-crossing action sequence to overcome obstacles, according to the obstacle avoidance strategy, includes the following steps: The real-time radar image and current movable options are input into the trained obstacle avoidance strategy, which outputs action commands. According to the action instructions, execute the preset obstacle-crossing action procedure; Step S200: In the non-obstacle-crossing state, control the robot to execute the automatic cleaning process, which includes lowering the cleaning mechanism and driving it to perform a linear cleaning motion along the curtain wall surface. Step S300: When it is necessary to change the cleaning direction, control the robot to perform the turning process.

[0051] In step S100, when the robot detects a traversable obstacle ahead, the following steps are included: S101. When the robot is in automatic mode and there is no uncleaned area behind it, it detects obstacles in front of it through the vision system. S102. If an obstacle is detected, obtain the distance and type parameters of the obstacle; S103. Determine whether it is necessary to cross the obstacle based on the distance to the obstacle, and determine whether it is possible to cross the obstacle based on the type of obstacle; S104. If the judgment result is that it is necessary and can be crossed, then it is determined that there is a crossable obstacle, and the preset obstacle crossing action process is executed. In some embodiments, step S104 involves executing a preset obstacle-crossing action procedure, including the following steps: S1041, The first adsorption mechanism located in the middle is released from adsorption and retracted; S1042. The forward and backward movement motors controlling the movement of the robot body stop outputting torque and switch to follow-up mode. S1043, The pull rope motors on both sides rotate forward, pulling the first adsorption mechanism forward; S1044. When the first adsorption mechanism moves to the positive limit signal that triggers the forward and backward moving motors, the rope motor stops. S1045. The suction cup of the first adsorption mechanism extends and adsorbs onto the surface of the curtain wall, forming a new stable support point. S1046, The suction cup of the second adsorption mechanism releases adsorption and retracts; S1047: The front and rear moving motors switch to torque mode and reverse to lift the retracted second adsorption mechanism upwards. S1048. When the front suction cup 21 of the second adsorption mechanism is raised to a height sufficient to overcome the obstacle, the forward and backward movement motor stops. S1049, The second adsorption mechanism extends and re-adsorbs onto the curtain wall surface; S10410, The first adsorption mechanism releases adsorption and is retracted; S10411, The front and rear moving motors switch to follow-up mode again; S10412, The two side rope motors rotate forward, pulling the first adsorption mechanism to continue moving forward; Based on the positive limit signals of the front and rear moving motors, determine whether there is sufficient space for the first adsorption mechanism to move forward: If sufficient, the rope-pulling motor will stop after the first adsorption mechanism has completely passed the obstacle, and the two suction cups of the first adsorption mechanism will extend and adhere to the curtain wall, thus completing the obstacle-crossing action; If insufficient, the rope-pulling motor stops when the first adsorption mechanism moves to the edge of the obstacle, and the lifting and crossing actions of steps S1041 to S10412 are repeated to ensure complete obstacle crossing.

[0052] The radar chart and movable options are input into the AI ​​deep reinforcement learning model; the AI ​​deep reinforcement learning model outputs action commands; the control program executes the corresponding preset obstacle-crossing action process according to the action commands; wherein, the AI ​​deep reinforcement learning model can be obtained through the model training methods of the above embodiments; The suction cups of the two middle first adsorption mechanisms 1 are released from adsorption and retracted; the front and rear moving motors 4 stop torque output and become follow-up, and the pull rope motors 3 on both sides rotate forward to drive the first adsorption mechanism 1 to move forward; after reaching the positive limit of the front and rear moving motors 4, the pull rope motors 3 stop, and the suction cups of the two middle first adsorption mechanisms 1 extend and adsorb; the front suction cup 21 and the rear suction cup 23 of the second adsorption mechanism 2 are released from adsorption and retracted. The front and rear moving motor 4 outputs torque mode and reverses, moving the front suction cup 21 and the rear suction cup 23 upward; after the front suction cup 21 has crossed the obstacle to the maximum extent, the front and rear moving motor 4 stops, and the front suction cup 21 and the rear suction cup 23 extend and adsorb; the suction cups of the two first adsorption mechanisms 1 release the adsorption state and retract. The front and rear moving motor 4 stops outputting torque and becomes a follow-up motor. The two side pull rope motors 3 rotate forward, driving the suction cups of the first adsorption mechanism 1 to move forward. It is determined whether the positive limit position of the front and rear moving motor 4 is sufficient for the suction cups of the two middle first adsorption mechanisms 1 to cross the obstacle. If so, the pull rope motor 3 stops after the suction cups of the two middle first adsorption mechanisms 1 cross the obstacle, and the suction cups of the two middle first adsorption mechanisms 1 extend and adsorb, completing the obstacle crossing action. If not, the pull rope motor 3 stops after the suction cups of the two middle first adsorption mechanisms 1 reach the position in front of the obstacle, and the movement action is repeated once to complete the obstacle crossing.

[0053] In some embodiments, step S200, the automatic cleaning process includes the following steps: S201. Keep the robot's first adsorption mechanism 1 adsorbing and retract the suction cup of the second adsorption mechanism 2. S202, control the cleaning U-axis rotation motor 6 and the rear suction cup lifting motor 24 to lower the cleaning mechanism 8 to the surface of the curtain wall; S203. Start the cleaning mechanism 8, and according to the decision command, resolve the motion target at the end of the cleaning mechanism 8 into control commands for multiple motors to drive the cleaning mechanism 8 to perform linear cleaning.

[0054] In some embodiments, step S300, the turning process includes the following steps: S301, retract the rear suction cup 23 of the second adsorption mechanism 2, and control the forward and backward moving motor 4 to rotate forward so that the second adsorption mechanism 2 moves downward; S302, Control the U-axis rotary motor 6 to rotate 90° clockwise to level the cleaning mechanism; S303, Control the central rotary motor 5 to rotate 180° to exchange the positions of the front suction cup 21 and the rear suction cup 23 of the second adsorption mechanism 2; S304. Control the cleaning U-axis rotation motor 6 to reverse 90°, so that the replaced rear suction cup 23 returns to its working position.

[0055] In some embodiments, an obstacle avoidance cleaning method for a curtain wall cleaning robot further includes: The process before obstacle avoidance and movement: The robot is in automatic mode. It determines whether there is an uncleaned area behind the robot. If not, it visually detects whether there is an obstacle. If there is an obstacle, it visually detects the obstacle parameters. The control program determines the current coordinates (XY) of the device based on the position of each motor. It determines whether the obstacle distance is sufficient to overcome the obstacle. It visually detects whether the obstacle type can be crossed. If so, it enters the obstacle-crossing state. Cleaning process: The suction cups of the first adsorption mechanism 1 in the middle maintain adsorption, while the front suction cup 21 and the rear suction cup 23 release adsorption and retract; the cleaning U-axis rotary motor 6 rotates 180° and, in conjunction with the rear suction cup lifting motor 24, lowers the cleaning mechanism 8; the cleaning mechanism 8 is started, and the movement direction of the cleaning mechanism 8 is specified by the AI ​​deep reinforcement learning model, and the target point and attitude of the end of the cleaning mechanism 8 at the next time moment are back-solved into the target angles of the front and rear moving motors 4, the central rotary motor 5, and the cleaning W-axis rotary motor 6 to achieve linear movement at the end of the cleaning process; Turning-around process: When the robot is in the turning-around state, the rear suction cup 23 is released from the suction state and retracted. The front and rear moving motors 4 output torque mode and rotate forward, moving the front suction cup 21 and the rear suction cup 23 downward. The cleaning U-axis rotary motor 6 rotates 90° forward to flatten the cleaning mechanism 8 and the rear suction cup 23. The central rotary motor 5 rotates 180° to swap the positions of the front suction cup 21 and the rear suction cup 23. The cleaning U-axis rotary motor 6 rotates 90° backward to make the rear suction cup 23 parallel to the curtain wall glass surface, completing the turning-around action.

[0056] The obstacle avoidance cleaning method of the curtain wall cleaning robot of the present invention has the following advantages: First, it enables the curtain wall cleaning robot to make autonomous decision-making to overcome obstacles, and through an end-to-end intelligent algorithm, it can autonomously and efficiently complete obstacle avoidance and cleaning. Second, the obstacle-crossing algorithm based on visual imitation learning significantly improves the robot's intelligence level; Third, the robot's cleaning efficiency for glass curtain walls was improved by using an obstacle-crossing algorithm based on reinforcement learning.

[0057] The present invention discloses an obstacle avoidance cleaning method for a curtain wall cleaning robot, which can be deployed in a curtain wall cleaning robot. The curtain wall cleaning robot includes: a first adsorption mechanism 1, a second adsorption mechanism 2, a rope-pulling motor 3, a forward / backward moving motor 4, a central rotary motor 5, a cleaning U-axis rotary motor 6, a cleaning W-axis rotary motor 7, a cleaning mechanism 8, a frame 9, a guide member 10, and a connecting member 11. Two rope-pulling motors 3 are mounted opposite each other at both ends of the frame 9. The first adsorption mechanism 1 and the central rotary motor 5 are both mounted in the middle of the frame 9. The side of the connecting member 11 and the input of the central rotary motor 5 are connected. The output shaft is connected to allow the connector 11 to rotate; the front and rear moving motor 4 is mounted on the connector 11, and the side of the guide 10 is slidably connected to the connector 11. The output end of the front and rear moving motor 4 is connected to the side of the guide 10 for driving the guide 10 to move on the connector 11; the first suction mechanism 1 is provided with a first suction cup lifting motor 101 for driving the first suction cup 100. The first suction cup lifting motor 101 is mounted on the frame 9, and the first suction cup 100 is connected to the output end of the first suction cup lifting motor 101; the second suction mechanism 2 is provided with a motor for driving the front suction cup 100 to move. The front suction cup lifting motor 22 of the plate 21 and the rear suction cup lifting motor 24 for driving the rear suction cup 23; the front suction cup lifting motor 22 is installed at one end of the guide member 10 and drives the front suction cup 21 through the first guide post 221. One end of the first guide post 221 is connected to the front suction cup 21, and the side of the first guide post 221 is slidably connected to one end of the guide member 10. The output end of the front suction cup lifting motor 22 is drivenly connected to the side of the first guide post 221 to drive the first guide post 221 to move on one end of the guide member 10; the cleaning U-axis rotary motor 8 is connected to the other end of the guide member 10, cleaning... The output end of the U-axis rotary motor 8 is connected to a limiting member 81. The rear suction cup lifting motor 24 is mounted on the limiting member 81 and drives the rear suction cup 23 through the second guide post 241. One end of the second guide post 241 is connected to the rear suction cup 23, and the side of the second guide post 241 is slidably connected to the limiting member 81. The output end of the rear suction cup lifting motor 24 is drivenly connected to the side of the second guide post 241 to drive the second guide post 241 to move on the limiting member 81. The W-axis rotary motor 7 is mounted on the other end of the second guide post 241, and the cleaning mechanism 8 is connected to the output end of the W-axis rotary motor 7. The cleaning mechanism 8 can be an electric roller brush. During the operation of the curtain wall cleaning robot, two ropes need to be lowered from the roof. The output ends of the two rope-pulling motors 3 are respectively connected to the rope transmission so that the rope-pulling motors 3 can drive the frame 9 to rise or fall.

[0058] It should be noted that the curtain wall cleaning robot used in this application has been disclosed in the prior art (CN121080852A). This application mainly protects the obstacle avoidance method applied to the robot. Other specific structural driving relationships can be obtained by those skilled in the art based on the prior art, and will not be described in detail here.

[0059] The advantages of the robot obstacle avoidance method of the present invention are as follows: 1. Enhance three-dimensional spatial understanding: Existing technologies are mostly based on two-dimensional image processing, lacking depth information and making it difficult to accurately identify the shapes of complex obstacles.

[0060] This invention utilizes a binocular camera and the PointNet++ algorithm to extract 3D point cloud information, providing a comprehensive understanding of the three-dimensional structure of the environment.

[0061] 2. Enhanced independent decision-making ability: Existing technology relies on pre-programming and simple sensor feedback, and cannot make autonomous decisions in real time.

[0062] This invention enables robots to generate and execute obstacle-crossing action sequences in real time through diffusion model training, achieving autonomous decision-making for obstacle avoidance and obstacle crossing, and thus exhibiting greater adaptability.

[0063] 3. High precision in motion generation: Existing technologies mostly employ traditional control algorithms, resulting in low accuracy and flexibility in motion generation.

[0064] This invention trains a neural network using expert-managed data to generate more accurate and flexible robot motion sequences, enabling it to handle a variety of complex scenarios.

[0065] 4. Enhanced ability to adapt to complex environments: Existing technology: The obstacle avoidance effect is poor when facing protruding window frames, glass seams, broken glass and glass holes.

[0066] This invention enhances the robot's obstacle avoidance and obstacle crossing capabilities in complex environments through an end-to-end intelligent algorithm, significantly improving cleaning efficiency and safety.

[0067] 5. Closed-loop control system: Existing technology: lacks real-time adjustment capabilities and struggles to cope with dynamic changes in the environment.

[0068] This invention: By using real-time sensor data feedback to form a closed-loop control system, the robot can adjust its plans in real time according to environmental changes, thereby improving its autonomous adaptability and operational accuracy.

[0069] A cleaning robot includes a memory 12 and a processor 13. The memory 12 stores a computer program, and the processor 13 executes the computer program to implement the robot obstacle avoidance method disclosed in the above embodiments.

[0070] This application discloses a computer-readable storage medium storing a computer program, wherein the computer program causes a computer to execute the model training method or robot obstacle avoidance method disclosed in the above embodiments; wherein the processor may be a central processing unit (CPU).

[0071] This application also discloses an application publishing platform, which is used to publish computer program products. When the computer program products are run on a computer, the computer performs some or all of the steps of the methods described in the above method embodiments.

[0072] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Those skilled in the art should also recognize that the embodiments described in the specification are optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0073] In the various embodiments of this application, it should be understood that the sequence number of each process does not necessarily imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0074] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they can be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0075] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0076] If the integrated units described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-accessible memory. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several requests to cause a computer device (which can be a personal computer, server, or network device, specifically a processor in the computer device) to execute some or all of the steps of the methods described in the various embodiments of this application.

[0077] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0078] Based on the disclosure and teachings of the foregoing specification, those skilled in the art can make changes and modifications to the above embodiments. Therefore, the invention is not limited to the specific embodiments described above, and any obvious improvements, substitutions, or modifications made by those skilled in the art based on this invention are within the scope of protection of this invention. Furthermore, although some specific terms are used in this specification, these terms are only for convenience of explanation and do not constitute any limitation on the invention.

Claims

1. A model training method, characterized in that, Includes the following steps: Obtain the expert demonstration dataset; Training samples are constructed from the dataset, and the training samples include historical point cloud sequences, historical ontology state sequences, and expert action sequences; Preprocess the historical point cloud sequence, the expert action sequence, and the historical ontology state sequence; Noise is added to the expert action sequence to generate noisy actions; Based on the historical point cloud sequence, the historical ontology state sequence, and the noisy action, a noise prediction network is trained to predict the added noise. The parameters of the noise prediction network are optimized based on the difference between the predicted noise and the added noise.

2. The model training method as described in claim 1, characterized in that, The step of training a noise prediction network based on the historical point cloud sequence, the historical ontology state sequence, and the noisy actions to predict added noise includes the following steps: The historical point cloud sequence is input into the first encoder to extract visual feature vectors; The historical ontology state sequence is input into the second encoder to extract motion feature vectors; The noisy action, the intensity of the added noise, the visual feature vector, and the motion feature vector are input into a UNet network to predict the noise added to the expert action sequence.

3. The model training method as described in claim 2, characterized in that, The step of optimizing the parameters of the noise prediction network based on the difference between the predicted noise and the added noise includes the following steps: Calculate the mean square error between the predicted noise output by the UNet network and the added noise; Based on the mean square error, the weights of the first encoder, the second encoder, and the UNet network are updated using backpropagation and gradient descent algorithms.

4. A robot obstacle avoidance method, characterized in that, Includes the following steps: Acquire data on tasks performed by robots operated by human experts, and generate expert demonstration datasets; Based on the expert demonstration dataset, determine the obstacle avoidance strategy; The robot obtains the current point cloud sequence and the current body state sequence of its environment, inputs the obstacle avoidance strategy, and generates a future action sequence.

5. The robot obstacle avoidance method as described in claim 4, characterized in that, The process of acquiring data from tasks performed by robots operated by human experts includes the following steps: In a virtual simulation environment, a human operator remotely controls the trainer aircraft from the trainer's perception perspective and captures motion data of each joint of the trainer aircraft during the execution of a task.

6. The robot obstacle avoidance method as described in claim 5, characterized in that, The process of generating the expert demonstration dataset includes the following steps: The captured joint motion data of the trainer machine is mapped onto the robot using a kinematic algorithm; The robot is controlled to perform the mapped actions, and sensor data and body status data of the robot are collected simultaneously. The synchronous data collected from multiple tasks are processed to generate an expert demonstration dataset; wherein, the expert demonstration dataset includes historical point cloud sequences, historical ontology state sequences, and expert action sequences.

7. The robot obstacle avoidance method as described in claim 6, characterized in that, The process of acquiring the current point cloud sequence and current robot state sequence of the robot's environment, and generating a future action sequence based on the obstacle avoidance strategy, includes the following steps: The robot acquires the real-time point cloud of the current environment and reads the current joint status; Based on current observations, the future action sequence is generated iteratively using the trained obstacle avoidance strategy.

8. The robot obstacle avoidance method as described in claim 4, characterized in that, After generating the future action sequence, the method further includes: Obtain the action of the first time step of the future action sequence and send it to the underlying controller for execution; After the robot completes one step, it reads new sensor data and repeats the process of generating future action sequences based on the new sensor data.

9. A method for obstacle avoidance cleaning using a curtain wall cleaning robot, characterized in that, Includes the following steps: When the robot detects a crossable obstacle in front of it, it is controlled to execute a preset obstacle-crossing action procedure to cross the obstacle according to the obstacle avoidance strategy. In a non-obstacle-crossing state, the robot is controlled to perform an automatic cleaning process, which includes lowering the cleaning mechanism and driving it to perform cleaning movements along the curtain wall surface. When it is necessary to change the cleaning direction, control the robot to perform a turning process.

10. A cleaning robot, comprising a memory and a processor, including a memory storing executable program code and a processor coupled to the memory; wherein, The processor calls the executable program code stored in the memory to execute the method as described in any one of claims 4 to 8.

Images