A world model based robot teleoperation prediction method, device and medium

Abstract

The present application relates to the technical field of robot operation, in particular to a robot teleoperation prediction method based on a world model, a device and a medium, the method comprising: based on the acquired current position and motion state of the robot, dynamically selecting at least one target camera associated with the current work area of the robot, acquiring real-time monitoring video data collected by the target camera and compressing the video data, uploading the compressed video data stream, the current position and motion state of the robot to the world model preset in the cloud, and after reconstructing the corresponding current world state, combining the current position and motion state of the robot to generate a next-time world state prediction result and issue it to the client, so that the operator generates an operation instruction to control the execution of the robot, and repeats the above steps to update the prediction of the next-time world state; the present application can reduce the influence of the operator's perception delay, and improve the stability and safety of the robot teleoperation.

Description

Technical Field

[0001] This invention relates to the field of robot operation technology, and in particular to a method, device and medium for predicting robot teleoperation based on a world model. Background Technology

[0002] With the rapid development of robotics technology, teleoperation systems have been widely used in fields such as hazardous environment operations, telemedicine, and space exploration. In existing robot teleoperation systems, operators typically rely on real-time monitoring video to control the robot, observing the scene and issuing operating commands to achieve remote control of the robot.

[0003] However, existing teleoperation methods have the following problems: First, communication and processing delays cause image lag. Video transmission inevitably produces end-to-end delays, causing the image seen by the operator to lag significantly behind the actual state, forming a perception delay bias, which affects the stability and smoothness of teleoperation. Moreover, when delays accumulate or the network fluctuates, teleoperation is prone to vibration and misoperation. Second, the difficulty in predicting operations leads to safety risks. In contact operations or dynamic environments, operators find it difficult to judge the consequences of operations based on the delayed image, which can easily lead to safety accidents such as collisions and instability.

[0004] In scenarios where similar problems exist, such as those encountered in online FPS games, client-side prediction and frame synchronization mechanisms can alleviate the issue, providing players with a smooth, low-latency interactive experience. However, games are built upon deterministic logic, which cannot be directly transferred to real-world robotic operation scenarios characterized by high complexity and non-determinism. Therefore, there is an urgent need for a teleoperation solution that combines real-world scenarios to improve prediction accuracy and uses the prediction results to assist human operation. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a robot teleoperation prediction method, device, and medium based on a world model. This method can utilize a world model to generate a short-term next-moment world state based on real-time monitoring video and robot state information, thereby reducing the impact of operator perception delay and improving the real-time performance, stability, and safety of robot teleoperation.

[0006] According to a first aspect of the present invention, a method for predicting robot teleoperation based on a world model is provided, comprising the following steps:

[0007] S1, based on the robot's current position and motion state, dynamically select at least one target camera associated with the robot's current working area from a number of monitoring cameras distributed in the working environment.

[0008] S2 acquires and compresses real-time monitoring video data collected by the target camera, and uploads the compressed video data stream, the robot's current position and motion status to a preset world model in the cloud.

[0009] S3, control the world model to reconstruct the corresponding current world state based on the compressed video data stream, and based on the current world state, the robot's current position and motion state, predict the world state at the next moment within a preset future time window, and generate the prediction result of the world state at the next moment.

[0010] S4, the next moment world state prediction result is sent to the client and presented in a perceptual form, so that the operator can generate operation instructions based on the presented next moment world state prediction result and send them to the robot for execution.

[0011] S5. Repeat steps S1 to S4 to continuously update the prediction of the world state at the next moment, forming a real-time closed-loop prediction of robot teleoperation.

[0012] According to a second aspect of the present invention, a non-transitory computer-readable storage medium is provided, wherein at least one instruction or at least one program is stored therein, the at least one instruction or the at least one program being loaded and executed by a processor to implement the above-described world model-based robot teleoperation prediction method.

[0013] According to a third aspect of the present invention, an electronic device is provided, including a processor and the aforementioned non-transitory computer-readable storage medium.

[0014] The present invention has at least the following beneficial effects:

[0015] This invention provides a robot teleoperation prediction method based on a world model. First, based on the robot's current position and motion state, at least one target camera associated with the robot's current working area is dynamically selected, effectively avoiding resource waste caused by full video data transmission. This ensures the world model can focus on the most critical environmental information for prediction and world state reconstruction. Then, real-time monitoring video data collected by the target camera is acquired and compressed. The compressed video data stream, the robot's current position, and motion state are uploaded to a pre-set world model in the cloud. After the world model reconstructs the corresponding current world state, it is combined with the robot's current position and motion state to generate a prediction result for the next moment's world state, which is then sent to the client. This allows the operator to generate operation commands to control the robot's execution. Simultaneously, the prediction of the next moment's world state is continuously updated. By using the world model to generate short-term predictions of the future world in the cloud and presenting them intuitively to the operator, this fundamentally compensates for the perception delay caused by communication and processing, providing the operator with a visual basis ahead of reality for decision-making. This significantly improves the operational stability and real-time performance of teleoperation, while also effectively enhancing operational accuracy and safety in dynamic and complex scenarios. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart of a robot teleoperation prediction method based on a world model provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention provides a robot teleoperation prediction method based on a world model, such as... Figure 1 As shown, the method includes the following steps:

[0020] S1, based on the robot's current position and motion state, dynamically select at least one target camera associated with the robot's current working area from several monitoring cameras distributed in the working environment. In specific implementations, it may also be considered to select target cameras associated with the robot's potential interaction areas simultaneously.

[0021] Specifically, the robot's current location is obtained through a positioning system; the positioning system includes, but is not limited to, indoor positioning systems, visual positioning systems, lidar positioning systems, or combinations thereof.

[0022] Specifically, the robot's motion state includes, but is not limited to, the robot's current pose information, current direction of motion, current speed of motion, and current operating state.

[0023] As described above, by dynamically selecting the monitoring camera most relevant to the robot, the waste of resources caused by the transmission of full video data is effectively avoided, thereby ensuring that the world model can focus on the most critical environmental information for prediction and world state reconstruction. Furthermore, the dynamic selection of monitoring cameras can switch perspectives in real time with the robot's movement, ensuring the operator's precise operation.

[0024] Specifically, the process of dynamically selecting a target camera includes:

[0025] S101: Obtain the robot's current position and motion state, and acquire the position information, field-of-view parameters, and effective imaging distance range of each monitoring camera. In specific implementation, the position information of the monitoring cameras is three-dimensional spatial coordinates, and the effective imaging distance range is obtained through the field-of-view parameters and orientation angles of the monitoring cameras.

[0026] Furthermore, based on the robot's current position, current direction of movement, and current speed, the robot's expected trajectory in the near future can be obtained, thus identifying target cameras associated with the robot's potential work area.

[0027] S102, for each surveillance camera, calculate the scores of three evaluation indicators: spatial proximity score, field of view coverage quality score, and task relevance score.

[0028] Specifically, the spatial proximity score is determined based on the distance between the robot's current position and the position of the monitoring camera, as well as the effective imaging range of the monitoring camera; the calculation formula is as follows:

[0029] Where S1 represents the spatial proximity score, and D i Let D be the Euclidean distance between the robot's current position and the position of the surveillance camera. maxThis represents the critical distance corresponding to the effective imaging range. In other words, when the robot's position is outside the effective imaging range, the spatial proximity score of the monitoring camera is 0; when it is within the effective imaging range, the closer the robot is to the monitoring camera, the higher the spatial proximity score.

[0030] Specifically, the field-of-view coverage quality score is determined based on whether the robot is within the field of view of the monitoring camera and the degree to which it deviates from the center of the field of view. The details are as follows:

[0031] If |α|≤FOV h / 2, and |β|≤FOV v / 2, then S2=(1-|α| / (FOV) h / 2))×(1-|β| / (FOV v / 2))×Q; otherwise, S2=0; where S2 represents the field of view coverage quality score, FOV h and FOV v These are the horizontal and vertical field of view, respectively. α is the horizontal angle relative to the direction the robot is facing relative to the monitoring camera, β is the vertical angle relative to the direction the robot is facing relative to the monitoring camera, and Q is the resolution and imaging quality coefficient of the monitoring camera obtained in advance.

[0032] In addition, for the robot's expected motion trajectory, the proportion of trajectory points falling within the camera's field of view is calculated as an additional predictive field of view coverage score.

[0033] Specifically, the task relevance score is determined based on the degree of matching between the current operation task type and the preset task relevance of the surveillance camera.

[0034] For example, if the current operation task is a fine operation task such as assembly or grasping, the monitoring camera that can provide a close-up view should be selected first. The task relevance score is inversely proportional to the distance from the monitoring camera to the robot's end effector.

[0035] If the current operation task is a safety monitoring task such as avoiding obstacles or preventing collisions, priority should be given to monitoring cameras that can cover the blind spots around the robot. The task relevance score is proportional to the additional field of view provided by the monitoring camera; where the additional field of view refers to the part that does not overlap with the field of view of other cameras.

[0036] If the current operation task type is a navigation and movement task such as going to the target area, the monitoring camera that can cover the robot's forward direction and the target area should be selected first. The task relevance score S3 = λ × F1 + (1-λ) × F2, where F1 is the coverage of the monitoring camera on the robot's forward path, F2 is the coverage of the monitoring camera on the target area, and λ is a preset weight coefficient.

[0037] S103, based on the robot's current movement speed and task operation scenario, dynamically assigns weight coefficients to each evaluation indicator and calculates a weighted average score for each monitoring camera. Specifically, this includes the following:

[0038] S1031, when V t <V min When, W2 > W1 > W3, where, V t V represents the robot's current speed. min To preset a low-speed threshold, W1, W2, and W3 are the weighting coefficients corresponding to S1, S2, and S3, respectively. When the robot is in a low-speed state, the importance of spatial proximity decreases due to the slow change in position. At this time, the operator needs to observe environmental details carefully, so the field of view coverage quality is assigned the highest weighting coefficient.

[0039] S1032, when V min ≤V t ≤V max When W1 = W2 > W3, where V max A preset medium-speed threshold is set. When the robot is moving at a medium speed, it needs to maintain both image clarity and the continuity of camera switching. Therefore, a balance is needed between proximity and field of view coverage quality to avoid frequent camera switching due to an excessive pursuit of image quality.

[0040] S1033, when V t >V max At that time, W1 > W2 > W3. When the robot is in high-speed mode, its position changes rapidly. If the camera is too far away, the video transmission delay will exacerbate the feeling of lag in operation. Therefore, cameras with high spatial proximity should be selected to shorten the video transmission path and reduce end-to-end latency.

[0041] S1034, When the task operation scenario is in a critical stage of the task, W1 < W2 < W3. When the robot approaches the target or performs delicate operations, the general viewpoint cannot meet the requirements. Therefore, the weight coefficient of task relevance is significantly increased to ensure that a dedicated camera that is aligned with the target area or can provide a critical operation viewpoint is selected.

[0042] S104. Select at least one surveillance camera as the target camera according to the overall score from high to low.

[0043] The above-mentioned method achieves precise selection of the optimal observation angle by selecting multi-dimensional evaluation indicators and calculating the comprehensive score of the surveillance camera, thus ensuring high-value surveillance video data. At the same time, the indicator weights are dynamically adjusted according to the movement speed and task scenario, so that the selection strategy can adapt to the needs of different operation stages, thereby improving the adaptability and robustness of the remote operating system.

[0044] S2 acquires and compresses real-time monitoring video data collected by the target camera, and uploads the compressed video data stream, the robot's current position and motion status to a preset world model in the cloud.

[0045] Specifically, the methods for compressing the real-time monitoring video data include:

[0046] Method 1: Directly compress and encode real-time monitoring video data or monitoring video segments obtained based on real-time monitoring video data as the compressed video data stream; this can be understood as: extracting key monitoring video segments from real-time monitoring video data for compression.

[0047] Method 2: Extract video features or perform embedding representation from real-time monitoring video data using a preset encoder, and use the video features or embedding representation as a compressed video data stream. For example, an embedding model can be used to convert images in real-time monitoring video data into embedded representations.

[0048] The compression processing method is adaptively selected based on the current network bandwidth, latency requirements, and / or cloud resource status.

[0049] As described above, by providing two compression methods—original video and feature embedding—the system can flexibly switch according to network bandwidth, latency requirements, and cloud resource status, effectively balancing video quality and transmission efficiency, improving latency, enhancing the system's robustness in complex network environments, and ensuring that it can provide stable, low-latency data input to the world model in different network environments, which is conducive to improving the prediction accuracy of the world model.

[0050] S3, control the world model to reconstruct the corresponding current world state based on the compressed video data stream, and based on the current world state, the robot's current position and motion state, predict the world state at the next moment within a preset future time window, and generate the prediction result of the world state at the next moment.

[0051] In practice, step S3 specifically includes the following:

[0052] S301, if the compressed video data stream is an embedded representation, then the embedded representation is decoded using a world model to reconstruct the video representation of the current world state.

[0053] S302, the video representation of the reconstructed current world state, combined with the robot's current position and motion state, is input into the world model as input information.

[0054] S303, the control world model, based on input information, generates a video frame sequence of the world state at the next moment within a preset future time window, state change trend data, or a combination of both, through an internal prediction network.

[0055] The above-mentioned method of using a world model to generate short-term predicted states of the future world in the cloud fundamentally compensates for the perception delay caused by communication and processing, enabling operators to know about environmental changes in advance, that is, to know the state of the world in the short term, turning passive reaction into active prediction. Moreover, based on highly accurate prediction results, it can significantly improve the operational stability and real-time performance of teleoperation.

[0056] Furthermore, in step S3, the preset future time window is dynamically adjusted through the following steps:

[0057] S310 acquires the robot's current movement speed, current task type, and prediction confidence of the world model in real time. The prediction confidence is inversely proportional to the prediction uncertainty score, which is estimated based on the blurriness of the predicted video frames, multi-frame consistency, or variance of the world model output.

[0058] S320 determines the base value of the corresponding time window based on the current operation task type. For example, the base value of the time window for a fine operation task is 100-200ms, with a short window to ensure prediction accuracy; the base value of the time window for a safety monitoring task is 200-400ms, with a medium window to balance timeliness and accuracy; and the base value of the time window for a navigation and movement task is 500-1000ms, with a long window to predict the path.

[0059] S330, perform a first-level correction on the base value of the time window based on the current motion speed; wherein, the length of the corrected time window is inversely proportional to the current motion speed. That is, extend the window at low speeds to predict the more distant future, maintain the base window at medium speeds, and shorten the window at high speeds to prioritize prediction reliability.

[0060] S340, a second-level correction is performed on the modified time window based on the prediction confidence of the world model; the length of the modified time window is proportional to the prediction confidence. That is, the lower the prediction confidence, the shorter the time window.

[0061] S350, the length of the time window after two levels of correction is used as the preset future time window.

[0062] The above-mentioned two-level dynamic correction of the time window by comprehensively considering the task type, movement speed and prediction confidence achieves adaptive adjustment of the prediction range. The window length is dynamically adjusted according to the prediction confidence. When the world model is uncertain about the prediction results, the window range is actively shortened to avoid presenting unreliable future information to the operator. This achieves the optimal balance between prediction timeliness and accuracy, significantly improving the stability of the remote operating system and the user experience.

[0063] S4, the next moment world state prediction result is sent to the client and presented in a perceptual form, so that the operator can generate operation instructions based on the presented next moment world state prediction result and send them to the robot for execution.

[0064] Specifically, the presentation format of the next moment world state prediction result includes: a standalone prediction video frame, a prediction trajectory or dynamic effect superimposed on the prediction video frame, and a graphical interface prompting the predicted risks or events that are about to occur. When superimposing the prediction trajectory and dynamic effects, a semi-transparent display method can be used.

[0065] The above-mentioned method presents the predicted future scene to the operator intuitively, providing the operator with a visual basis that is ahead of reality for decision-making. This allows the operator to operate based on future risks. When the operation command reaches the robot and is executed after transmission delay, its response time on the predicted scene is exactly synchronized with the actual environmental changes. This proactive prediction can effectively compensate for the time misalignment caused by communication delay, thereby effectively improving the safety and accuracy of operation in dynamic and complex scenarios.

[0066] In its specific implementation, the method further includes the following steps:

[0067] The predicted world state at the next moment is sent as a safety constraint signal to the robot's underlying control system so that deceleration or emergency stop operations are automatically triggered when a dangerous event is predicted to occur.

[0068] S5. Repeat steps S1 to S4 to continuously update the prediction of the world state at the next moment, forming a real-time closed-loop prediction of robot teleoperation.

[0069] Through continuous rolling predictions and updates, the system can always provide operators with the latest and most accurate predictions of future states. These predictions can present a continuous sequence of future video frames, providing operators with rich dynamic trend information, ensuring the real-time performance and stability of the entire teleoperation process. Furthermore, it can continuously correct subsequent predictions based on the actual execution results of the robot, forming a real-time closed-loop teleoperation, which significantly improves the success rate of long-term tasks.

[0070] In a preferred embodiment, the prediction module of the world model is constructed through the following steps:

[0071] S10: Acquire real monitoring video data or simulation rendering data corresponding to several robot operation scenarios to construct a training dataset. The training dataset includes the robot's position, state, and the current monitoring video. The current video frame and the current robot position and state are used as input, and a sequence of future video frames is used as the supervision target, thus forming a complete training dataset.

[0072] S20, using a pre-trained generative model as a teacher model, the teacher model is fine-tuned using the training dataset so that the teacher model can predict future video frames based on the current monitoring video data and the robot's current state; the pre-trained generative model is a diffusion model or a streaming model.

[0073] When using a diffusion model for prediction, based on the input data, a trained denoising diffusion probability model is used to iteratively denoise from random noise to generate future video frames that match the conditional input. To meet real-time requirements, an accelerated sampling algorithm can be used to reduce the number of iterations.

[0074] When using a streaming model for prediction, a standardized streaming model is constructed using a reversible neural network. Through a series of reversible transformations, a simple prior distribution is mapped to a complex future video frame distribution, enabling single forward inference prediction from conditional input to future video frames and meeting the low latency requirement.

[0075] S30: Build a lightweight student model and input the training dataset into the student model to generate the corresponding prediction results.

[0076] S40 uses the teacher model's prediction results on the training dataset as a supervision signal, and trains the student model using knowledge distillation techniques based on the prediction results generated by the student model. The trained student model is then deployed in the prediction module of the world model. This can be understood as: providing knowledge distillation techniques to train the student model is to make the prediction output of the student model close to the prediction output of the teacher model.

[0077] As described above, the future prediction capabilities of a large-scale pre-trained generative model are transferred to a lightweight student model through a teacher-student model distillation architecture. This significantly reduces the number of model parameters and inference latency while maintaining prediction accuracy, enabling the model to meet the millisecond-level response requirements of real-time teleoperation and laying the foundation for subsequent real-time prediction presentation.

[0078] Embodiments of the present invention also provide a non-transitory computer-readable storage medium that can be disposed in an electronic device to store at least one instruction or at least one program related to implementing a method in the method embodiments, wherein the at least one instruction or the at least one program is loaded and executed by the processor to implement the method provided in the above embodiments.

[0079] Embodiments of the present invention also provide an electronic device, including a processor and the aforementioned non-transitory computer-readable storage medium.

[0080] While specific embodiments of the invention have been described in detail by way of example, those skilled in the art should understand that the examples are for illustrative purposes only and not intended to limit the scope of the invention. It should also be understood that various modifications can be made to the embodiments without departing from the scope and spirit of the invention. The scope of the invention is defined by the appended claims.

Claims

1. A robot teleoperation prediction method based on a world model, characterized in that, The method includes the following steps: S1, based on the robot's current position and motion state, dynamically select at least one target camera associated with the robot's current working area from a number of monitoring cameras distributed in the working environment. S2, acquire and compress the real-time monitoring video data collected by the target camera, and upload the compressed video data stream, the robot's current position and motion state to the preset world model in the cloud; S3, control the world model to reconstruct the corresponding current world state based on the compressed video data stream, and predict the world state at the next moment within a preset future time window based on the current world state, the robot's current position and motion state, generating a prediction result for the world state at the next moment; wherein, the preset future time window is dynamically adjusted through the following steps: real-time acquisition of the robot's current motion speed, current operation task type, and the prediction confidence of the world model; determining the corresponding time window base value according to the current operation task type; performing a first-level correction on the time window base value according to the current motion speed; wherein, the length of the corrected time window is inversely proportional to the current motion speed; performing a second-level correction on the corrected time window according to the prediction confidence of the world model; wherein, the length of the corrected time window is directly proportional to the prediction confidence; and using the length of the time window after the two-level correction as the preset future time window; S4, the next moment world state prediction result is sent to the client and presented in a perception form, so that the operator can generate operation instructions based on the presented next moment world state prediction result and send them to the robot for execution; S5. Repeat steps S1 to S4 to continuously update the prediction of the world state at the next moment, forming a real-time closed-loop prediction of robot teleoperation.

2. The robot teleoperation prediction method based on a world model according to claim 1, characterized in that, In step S1, the process of dynamically selecting the target camera includes: S101, obtain the robot's current position and motion state, and obtain the position information, field of view parameters and effective imaging distance range of each monitoring camera; S102, for each surveillance camera, calculate the scores for three evaluation indicators: spatial proximity score, field of view coverage quality score, and task relevance score. The spatial proximity score is determined based on the distance between the robot's current position and the surveillance camera's position, as well as the effective imaging distance range of the surveillance camera. The field of view coverage quality score is determined based on whether the robot is within the surveillance camera's field of view and the degree to which it deviates from the center of the field of view. The task relevance score is determined based on the degree of matching between the current operation task type and the surveillance camera's preset task relevance. S103: Based on the robot's current movement speed and task operation scenario, dynamically allocate the weight coefficients of each evaluation index and calculate the comprehensive score of each monitoring camera by weighting. S104. Select at least one surveillance camera as the target camera according to the overall score from high to low.

3. The robot teleoperation prediction method based on a world model according to claim 1, characterized in that, In step S2, the compression processing of the real-time monitoring video data includes: The real-time monitoring video data or monitoring video clips obtained based on the real-time monitoring video data are directly compressed and encoded to form the compressed video data stream; or A preset encoder is used to extract video features or embed them into real-time monitoring video data, and the video features or embedding representations are used as compressed video data streams. The compression processing method is adaptively selected based on the current network bandwidth, latency requirements, and / or cloud resource status.

4. The robot teleoperation prediction method based on a world model according to claim 1, characterized in that, Step S3 includes: S301, if the compressed video data stream is an embedded representation, then the embedded representation is decoded using a world model to reconstruct the video representation of the current world state; S302, the video representation of the reconstructed current world state, combined with the robot's current position and motion state, is input into the world model as input information; S303, the control world model, based on input information, generates a video frame sequence of the world state at the next moment within a preset future time window, state change trend data, or a combination of both, through an internal prediction network.

5. The robot teleoperation prediction method based on a world model according to claim 1, characterized in that, The presentation of the next moment world state prediction results includes: independent prediction video footage, prediction trajectories or dynamic effects superimposed on the prediction video footage, and predictions of impending risks or events displayed through a graphical interface.

6. The robot teleoperation prediction method based on a world model according to claim 1, characterized in that, The prediction module of the world model is constructed using the following steps: S10: Obtain real monitoring video data or simulation rendering data corresponding to several robot operation scenarios, and construct a training dataset. S20, using a pre-trained generative model as a teacher model, the teacher model is fine-tuned using the training dataset so that the teacher model can predict future video frames based on the current monitoring video data and the robot's current state; the pre-trained generative model is a diffusion model or a streaming model. S30: Build a lightweight student model and input the training dataset into the student model to generate the corresponding prediction results; S40 uses the teacher model's prediction results on the training dataset as a supervision signal, and trains the student model using knowledge distillation techniques based on the prediction results generated by the student model. The trained student model is then deployed in the prediction module of the world model.

7. The robot teleoperation prediction method based on a world model according to claim 1, characterized in that, The method further includes the following steps: The predicted world state at the next moment is sent as a safety constraint signal to the robot's underlying control system so that deceleration or emergency stop operations are automatically triggered when a dangerous event is predicted to occur.

8. A non-transitory computer-readable storage medium, wherein the storage medium stores at least one instruction or at least one program segment, characterized in that, The at least one instruction or the at least one program segment is loaded and executed by the processor to implement the world model-based robot teleoperation prediction method as described in any one of claims 1-7.

9. An electronic device, characterized in that, Includes a processor and the non-transitory computer-readable storage medium as described in claim 8.

Images