A robot, a trajectory correction method thereof and a device for training embodied intelligence model correction capability

Abstract

This application discloses a robot, its trajectory correction method, and a training device for the correction capability of an embodied intelligent model. The trajectory correction method includes: acquiring perception data and task instructions after the robot makes an execution error; inputting the perception data and task instructions into a pre-trained VLA model with trajectory correction capability; wherein the VLA model is obtained by learning a correction trajectory; processing the perception data and task instructions using the VLA model to generate a correction action sequence; controlling the robot to sequentially execute multiple correction actions in the correction action sequence, so that the robot's actual trajectory gradually approaches and eventually returns to the standard trajectory. Applying this application can provide effective correction capabilities for embodied intelligent models and robots, improving the system robustness and practicality of embodied intelligent models and robots.

Description

Technical Field

[0001] This application relates to the fields of robot control and artificial intelligence technology, specifically to a robot, its trajectory correction method, and a training device for the correction capability of an embodied intelligence model. Background Technology

[0002] In recent years, embodied AI, as an important branch of artificial intelligence, aims to enable intelligent agents to complete complex tasks through perception, decision-making, and action in physical environments. In these tasks, the precise execution of predefined trajectories (i.e., standard trajectories) by robots is a fundamental capability for many applications such as industrial automation, service robots, and autonomous driving. However, in actual deployments, due to factors such as sensor noise, model inaccuracies, dynamic environmental changes (e.g., object movement, changes in lighting), and unforeseen physical disturbances (e.g., collisions, slippage), robots often exhibit trajectory deviations when performing tasks, leading to task failures and even safety issues.

[0003] With the development of deep learning technology, methods based on vision-language-action (VLA) models have provided new possibilities for robot control. These models can directly learn the mapping from multimodal perception (such as images and language commands) to actions end-to-end, demonstrating strong generalization ability and adaptability to complex environments. However, current VLA models often lack effective error correction capabilities when faced with execution deviations, affecting the robustness and practicality of embodied intelligence model systems. Summary of the Invention

[0004] This application provides a robot and its trajectory correction method, as well as a training device for the correction capability of an embodied intelligent model, which can improve the system robustness and practicality of the embodied intelligent model and the robot.

[0005] To achieve the above objectives, this application adopts the following technical solution:

[0006] A trajectory correction method for a robot includes:

[0007] Acquire perception data and task instructions after the robot encounters an execution error;

[0008] The perceived data and task instructions are input into a pre-trained VLA model with trajectory correction capabilities; wherein the VLA model is obtained by learning the correction trajectory.

[0009] The VLA model is used to process the perceived data and task instructions to generate a sequence of corrective actions;

[0010] The robot is controlled to execute multiple correction actions in the correction action sequence in sequence, so that the robot's actual trajectory gradually approaches and eventually returns to the standard trajectory.

[0011] Preferably, the training process of the VLA model includes:

[0012] Acquire correction demonstration data; the correction demonstration data includes trajectory data of the robot starting to perform correction actions after an execution error occurs until it returns to the standard trajectory;

[0013] From the correction demonstration data, the correction trajectory is extracted; the starting state of the correction trajectory is the starting state for executing the correction action, and the state sequence contained in the correction trajectory is used to characterize the complete correction process that enables the robot to return from the error state to the standard trajectory.

[0014] First training data for training the correction capability is generated based on the correction trajectory, and the first training data is added to the training dataset of the VLA model.

[0015] A VLA model with trajectory correction capabilities is generated based on the training dataset.

[0016] Preferably, the interception and correction trajectory includes:

[0017] Determine the starting point of the correction action in the correction demonstration data;

[0018] The trajectory segment from the starting point of the correction action until the robot's action sequence coincides with the corresponding action sequence in the standard trajectory or reaches the preset endpoint is extracted as the correction trajectory.

[0019] Preferably, the step of generating first training data for training the correction capability based on the correction trajectory includes:

[0020] Training samples are obtained from the correction trajectory to form the first training data;

[0021] Wherein, at least one input state of the training sample contains first state information in the correction trajectory used to characterize the error state that needs to be corrected.

[0022] Preferably, the correction demonstration data includes in-situ correction demonstration data and / or rollback correction demonstration data;

[0023] The on-site correction demonstration data refers to the demonstration data of the robot planning a smooth trajectory to return to the standard trajectory and the next key point based on the current error state after an execution error occurs; the key point is the position point that must be reached in the standard trajectory.

[0024] The backtracking and correction demonstration data is a demonstration of how the robot, after an execution error occurs, retreats to the position on the standard trajectory before the execution error occurred, starting from the execution of the correction action.

[0025] Preferably, before acquiring the correction demonstration data, the method further includes: if the state deviation between the initial error state after the erroneous operation ends and the next key point on the standard trajectory is less than or equal to a preset threshold, then it is determined that an on-site correction strategy is adopted; if the state deviation is greater than the preset threshold, then it is determined that a backtracking correction strategy is adopted.

[0026] The acquisition of correction demonstration data includes: collecting corresponding correction demonstration data based on a determined correction strategy.

[0027] Preferably, before determining the relationship between the state deviation and the preset threshold, the method further includes:

[0028] If the initial error state can directly regress to the next key point of the standard trajectory, then the relationship between the state deviation and the preset threshold is determined, and an on-site correction strategy is adopted; if the initial error state cannot directly regress to the next key point of the standard trajectory, then a backtracking correction strategy is adopted.

[0029] A robot includes: a perception module, a decision-making and control module, and an execution module;

[0030] The perception module is used to acquire perception data in real time and send the acquired perception data and natural language commands to the decision and control module.

[0031] The decision and control module is used to input the acquired perception data and natural language instructions into a pre-trained VLA model with trajectory correction capabilities after the robot makes an execution error. The VLA model then processes the data to generate a sequence of corrective actions. The VLA model is obtained by learning the correction trajectory.

[0032] The execution module is used to drive the robot's end effector to sequentially execute multiple correction actions in the correction action sequence, so that the robot's actual trajectory gradually approaches and eventually returns to the standard trajectory.

[0033] A training device for the correction capability of an embodied intelligent model includes: a correction demonstration data acquisition unit, a correction trajectory interception unit, a training data generation unit, and a training unit.

[0034] The correction demonstration data acquisition unit is used to acquire correction demonstration data; wherein, the correction demonstration data includes trajectory data of the robot performing correction actions after an execution error occurs until it returns to the standard trajectory;

[0035] The correction trajectory extraction unit is used to extract the correction trajectory from the correction demonstration data; wherein the starting state of the correction trajectory begins with the execution error, and the state sequence contained in the correction trajectory is used to characterize the complete correction process that enables the robot to return from the error state to the standard trajectory.

[0036] The training data generation unit is used to generate first training data for training the correction capability based on the correction trajectory, and to add the first training data to the training dataset of the VLA model.

[0037] The training unit is used to train and generate a VLA model with trajectory correction capability based on the training dataset.

[0038] A computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, can implement the trajectory correction method for a robot as described above.

[0039] An electronic device, comprising at least a computer-readable storage medium and a processor;

[0040] The processor is configured to read the executable instructions from the computer-readable storage medium and execute the instructions to implement the trajectory correction method for the robot as described in any of the preceding claims.

[0041] As can be seen from the above technical solution, this application first acquires the perception data and task instructions after the robot makes an execution error, and inputs the perception data and task instructions into a pre-trained VLA model with trajectory correction capabilities. The VLA model is obtained by learning a correction trajectory. Then, the VLA model processes the perception data and task instructions to generate a sequence of correction actions. The robot is then controlled to execute multiple correction actions in the sequence of correction actions sequentially, causing the robot's actual trajectory to gradually approach and eventually return to the standard trajectory. In the above process, the VLA model is generated by pre-learning a correction trajectory and possesses trajectory correction capabilities. Therefore, it can provide the robot with effective correction capabilities. Based on this VLA model, multiple correction actions can be generated to enable the robot to return to the correct trajectory, thereby effectively improving the robustness and practicality of the embodied intelligence model and the robot system. Attached Figure Description

[0042] Figure 1 This is a schematic diagram illustrating the basic process of the trajectory correction method for the robot in this application;

[0043] Figure 2 This is a schematic diagram of the basic process of training the embodied intelligence model's error correction capability in this application;

[0044] Figure 3This is a schematic diagram illustrating the specific process of the training method for the correction capability of the embodied intelligent model and the trajectory correction method in a specific embodiment of this application.

[0045] Figure 4 This is a schematic diagram illustrating the processing using a composite correction strategy in a specific embodiment of this application;

[0046] Figure 5a and Figure 5b These are schematic diagrams of the trajectory of the on-site correction strategy and the backoff correction strategy in specific embodiments of this application, respectively.

[0047] Figure 6 This is a schematic diagram of the basic structure of the robot in this application;

[0048] Figure 7 A schematic diagram of the basic structure of a training device for corrective capabilities of an embodied intelligent model;

[0049] Figure 8 This is a schematic diagram of the basic structure of the electronic device provided in this application. Detailed Implementation

[0050] To make the objectives, technical means, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings.

[0051] This invention provides a method for trajectory correction of a robot and a method for training the correction capability of an embodied intelligence model. The core of this method is to use a VLA model generated from the robot's actual correction process data to process the perception data and task instructions after the robot's actions deviate, generate a correction action sequence, and enable the robot to correct the wrong actions and return to the correct standard trajectory.

[0052] In the VLA model training, a high-quality training dataset is constructed based on real trajectory correction process data, thereby giving the VLA model autonomous trajectory correction capabilities.

[0053] Figure 1 This is a schematic diagram illustrating the basic process of the trajectory correction method for the robot in this application. Figure 1 As shown, the method mainly includes the following steps:

[0054] Step 101: Obtain the perception data and task instructions after the robot encounters an execution error.

[0055] After a deviation in the robot's movement, the system acquires current perception data and user-issued task commands. These commands may have been issued before the deviation occurred or may have been issued again after the deviation. The acquired perception data includes robot perception data and environmental perception data.

[0056] Step 102: Input the perception data and task instructions into the pre-trained VLA model with trajectory correction capabilities.

[0057] The VLA model is obtained by learning from the correction trajectory, thus possessing trajectory correction capabilities. The correction trajectory is used to mark the complete correction process that allows the robot to return from an erroneous state to a standard trajectory.

[0058] Step 103: Use the VLA model to process the perception data and task instructions to generate a sequence of corrective actions.

[0059] Because the VLA model possesses trajectory correction capabilities, it can be used to process perceived data and task commands to generate a sequence of corrective actions. This sequence can be used to correct erroneous robot movements, guiding them back to the correct standard trajectory. Here, the standard trajectory refers to the trajectory corresponding to the robot's correct execution of the commanded actions.

[0060] Step 104: Control the robot to execute multiple correction actions in the correction action sequence in sequence, so that the robot's actual trajectory gradually approaches and eventually returns to the standard trajectory.

[0061] The robot executes a sequence of corrective actions to return to the standard trajectory, thus completing the trajectory correction.

[0062] At this point, Figure 1 The illustrated method flow concludes here. Based on a mature VLA model trained from the correction trajectory, this application provides a trajectory correction method for robots, applicable to online correction scenarios. Specifically, when the robot detects a deviation from its expected trajectory, it inputs its current multimodal state (such as joint angles, end-effector pose, camera images, etc.) into the trained VLA model in real time. The model then outputs a correction action sequence, controlling the robot to gradually correct the deviation until it returns to the correct task trajectory. This allows the robot system to maintain a high task completion rate and autonomy in dynamic and uncertain environments, significantly reducing reliance on human intervention. Therefore, by employing the trajectory correction method of this application, the robot's error recovery capability can be improved, enabling it to adjust its state from errors, replan its path, and return to the standard trajectory, thus ensuring stable task completion when encountering interference or deviations.

[0063] In the above Figure 1 The robot trajectory correction method shown uses a pre-trained VLA model to generate correction actions. This application further provides a training method for this VLA model with correction capabilities to generate the corresponding VLA model. Figure 2 The method for training the corrective capabilities of the embodied intelligence model provided in this application, such as... Figure 2 As shown, the method includes:

[0064] Step 201: Obtain correction demonstration data.

[0065] This step aims to collect complete process data on the robot's execution of a task, including the initiation of corrective actions after a deviation occurs and the successful correction. Specifically, the correction demonstration data includes at least the trajectory data from the robot's commencement of corrective actions after the error until it returns to the standard trajectory. This correction demonstration data not only records the complete action sequence from the robot's commencement of corrective actions to its return to the standard trajectory after the error occurs, but may also include environmental perception information (such as visual images) at the corresponding moments. Because this data originates from correction behavior in a real physical environment, it ensures the authenticity of subsequent training data, laying a solid foundation for the model to learn correction strategies that adapt to the real world.

[0066] Step 202: Extract the correction trajectory from the correction demonstration data.

[0067] After obtaining the original correction demonstration data, it is necessary to accurately extract the segments that truly reflect the "correction" process, namely the correction trajectory.

[0068] Specifically, the initial state where the robot ends an erroneous action and begins to execute a corrective action is first identified, i.e., the starting point of the corrective action. Then, the complete sequence of actions from this starting point until the robot coincides with the standard trajectory or the task ends is extracted as the corrective trajectory. This step is crucial because it filters out erroneous operation data that leads to deviations, ensuring that the training data is highly focused on the active recovery process of "recovering from an erroneous state to a correct state," providing the model with clean and goal-oriented learning samples.

[0069] Step 203: Generate the first training data for training the correction capability based on the correction trajectory.

[0070] To fully utilize the limited correction trajectories and enhance the model's generalization ability, this step uses sampling techniques to generate multiple training samples from each correction trajectory to form the first training data. Each training sample includes input data and output data. The input data contains at least information representing the error state to be corrected in the correction trajectory (i.e., the initial state of the correction trajectory or an error state during the process), and the output data is the corresponding correction action or the desired next state. This construction method forces the model to establish a direct mapping between "error states" and "correction actions" during learning. By densely sampling different error states, the model can learn diverse recovery strategies to cope with various deviations.

[0071] Step 204: Add the first training data to the training dataset of the VLA model.

[0072] Among common embodied intelligence models, the VLA model possesses powerful multimodal understanding and generation capabilities. Based on this, the embodied intelligence model in this application adopts the VLA model, which can simultaneously process visual observation, language commands, and action sequences.

[0073] To ensure the trained model can perform both routine tasks and handle execution biases, this application merges the first training data generated in step 203, which focuses on bias correction, with a large amount of normal task execution trajectory data (i.e., standard trajectory) data to form the complete training dataset for the VLA model. This hybrid training strategy aims to prevent the model from compromising its original standard task execution capabilities due to excessive focus on bias correction. During training, the model will simultaneously learn "how to execute correctly" and "how to recover after errors," enabling the trained model to efficiently execute the main task under normal conditions and effectively correct biases when detected.

[0074] Step 205: Train and generate a VLA model with trajectory correction capability based on the training dataset.

[0075] Using the rich dataset containing correction samples constructed in step 204, the VLA model is trained end-to-end. The trained VLA model then possesses inherent trajectory correction capabilities: it can understand the task objective, identify the current state, and generate a series of action instructions to guide the system back to the correct trajectory.

[0076] At this point, Figure 2 The training method described above concludes here. In this method, a dedicated training dataset is constructed by collecting complete data on the robot's journey from the start of corrective actions to restoration to the standard trajectory after a deviation occurs during task execution. This dataset is then incorporated into the existing VLA training dataset to achieve end-to-end training of the VLA model. This enables the trained VLA model to possess trajectory correction capabilities, automatically predicting and regressing to the correct path when faced with execution errors, achieving autonomous and stable correction from the erroneous state to the standard trajectory. Furthermore, this training set contains only pure data from the correction phase, excluding erroneous trajectories, thus avoiding the learning of incorrect operations. This effectively improves training efficiency and the effectiveness of correction actions, enhancing the robustness and practicality of the embodied intelligence model and the robot system.

[0077] In the above Figure 2Building upon the training process shown, to further cover more comprehensive correction scenarios, the correction demonstration data is designed to include data containing two typical strategies: in-situ correction demonstration data and backtracking correction demonstration data. Preferably, in-situ correction is suitable for minor deviations. The robot directly plans a smooth path back to the next key point on the standard trajectory based on the current error state. Its advantages are high correction efficiency and a natural path. Backtracking correction is suitable for severe deviations or unpredictable environmental changes. The robot first safely backtracks to a state on the standard trajectory before the execution error occurred, thereby undoing the impact of the erroneous operation and returning the system to a known, restartable state. Then, it can replan and execute the task based on the latest environmental perception information. Its advantages are high safety and strong robustness. By collecting data for these two strategies for training, the model can learn to flexibly select or fuse the most suitable correction strategy according to the severity of the deviation and the environmental context. Key points refer to the positions that must be reached on the standard trajectory, which are usually predefined.

[0078] During the data acquisition phase, various error types and their corresponding correction strategies are typically predefined. Subsequently, a robot is manually controlled remotely to perform correction processing according to the preset strategies while simultaneously collecting data. This method enables the acquisition of more representative and targeted correction data, providing learning examples for the model's subsequent autonomous decision-making, thereby improving the rationality and security of the entire system's correction decisions.

[0079] The specific implementation of the trajectory correction method and the correction capability training method in this application is illustrated below through specific embodiments. Figure 3 This is a schematic diagram illustrating the specific flowcharts of the training method and trajectory correction method for the embodied intelligent model in a particular embodiment of this application. This embodiment provides a detailed description of the implementation of the two methods by covering the complete training and correction process. Figure 3 As shown, the corresponding method flow specifically includes:

[0080] Step 301: Strategically collect demonstration data for error correction.

[0081] Deploy robotic systems to ensure they have the ability to record complete trajectory data, including joint states, end-effector poses, timestamps, and synchronized environmental perception data streams (such as RGB-D images).

[0082] During the robot's execution of a preset standard task, different types of execution deviations (such as positional offsets and posture errors) are introduced. When a robot execution error occurs, a correction process begins until it returns to the standard trajectory, and correction demonstration data generated throughout the process is collected. Specific correction operations can be set according to actual conditions. This embodiment provides two correction strategies: an in-situ correction strategy and a backtracking correction strategy. Corresponding correction operations can be executed according to the pre-set correction strategies.

[0083] As mentioned earlier, the on-site correction strategy refers to directly planning a smooth path back to the next key point of the standard trajectory based on the current error state after an execution error occurs; the backtracking correction strategy refers to backtracking from the current error state to a state on the standard trajectory before the execution error occurred.

[0084] The pre-defined correction strategy can be either an on-site correction strategy or a rollback correction strategy. When an execution error occurs, the corresponding correction operation is executed according to one of the pre-defined correction strategies. Alternatively, the pre-defined correction strategy can also be a combination of both. For example, when condition one is met, the on-site correction strategy is used to execute the corresponding correction operation; when condition two is met, the rollback correction strategy is used to execute the corresponding correction operation.

[0085] This embodiment employs a composite correction strategy: an on-site correction strategy when the execution deviation is small, and a backtracking correction strategy when the execution deviation is large. Specific processing is as follows: Figure 4 As shown, after an erroneous operation occurs, it is determined whether the robot can directly return to the next key point of the standard trajectory from the current initial erroneous state. If not, a backtracking correction strategy is used to plan a path to perform the correction operation; if so, it is further determined whether the state deviation between the initial erroneous state and the next key point of the standard trajectory is less than or equal to a preset threshold. If the deviation is less than or equal to the preset threshold, it indicates that the robot is more likely to enter the next key point state of the standard trajectory, and an in-situ correction strategy is adopted to plan a path to the next key point to perform the correction operation; otherwise, it indicates that it is more feasible for the robot to return to the key point state before the error, and a backtracking correction strategy is adopted to plan a path to the reset point to perform the correction operation.

[0086] Furthermore, the degree of state deviation caused by execution errors varies depending on the nature of the task, thus requiring different corrective strategies. For example, placement tasks typically experience smaller state deviations after an execution error, so an in-situ corrective strategy can be used: such as... Figure 5aAs shown, if an object deviates from its intended target position during placement, the correction trajectory can start from the deviation point, instructing the robot to plan a corrective trajectory to move the object to the intended position. When an execution error occurs in a grasping task, the state deviation is usually significant; therefore, a backtracking correction strategy can be set up. Figure 5b As shown, when the gripper deviates from its gripping position, the robot can retract to the preset position before gripping and restart the gripping action, thus avoiding further deviation caused by continuing to operate in an incorrect state.

[0087] From the start of the corrective action, the system collects demonstration data, continuously recording all relevant robot status data, observed images, and corresponding task instructions until the robot confirms it has returned to the standard trajectory. The data recorded from the start of the corrective action to the return to the standard trajectory can be considered a single instance of corrective demonstration data.

[0088] Step 302: Extract the correction trajectory from the correction demonstration data.

[0089] In this embodiment, all data from the correction demonstration data, starting from the start of the correction action and ending at the standard trajectory, are extracted to obtain the correction trajectory.

[0090] The starting point of the correction trajectory is the time point at which the correction action begins. This time point can be determined manually or by comparing the actual trajectory of the correction demonstration data with the standard trajectory.

[0091] Specifically, manual triggering can be achieved by manually recording the start time of the correction action;

[0092] The trajectory comparison can be performed by aligning the actual trajectory in the correction demonstration data with a pre-stored standard trajectory in time and space. A preset algorithm automatically identifies the turning point where the actual trajectory and the standard trajectory significantly separate, and then converge towards the standard trajectory for the first time. This point is determined as the starting point of the correction action, and the corresponding time point is the start time of the correction action.

[0093] Starting from the defined starting point, subsequent correction demonstration data is extracted until the actual trajectory coincides with the standard trajectory (error less than the threshold) or the task terminates, forming a correction trajectory. This trajectory eliminates erroneous operation segments that cause deviations and only contains the pure correction and recovery process. This processing ensures that data from erroneous processes is avoided during subsequent VLA model training, thereby preventing the VLA model from learning incorrect operations.

[0094] Step 303: Generate the first training data for training the correction capability based on the correction trajectory, and construct the complete training dataset.

[0095] Training samples are obtained by sampling data from each correction trajectory. Specifically, one or more states can be selected as input state sets from the state sequence of the correction trajectory according to a first preset rule. At the same time, one or more states are selected as target state sets from the same correction trajectory based on the state positions corresponding to the input state sets according to a second preset rule.

[0096] The first and second preset rules can typically be the same, for example, selected consecutively or non-consecutively, such as selected at equal intervals or by key states. In fact, the input state set represents historical states and needs to include first-state information representing the erroneous states that need to be corrected; the target state set represents the desired future states. These states belong to the same correction trajectory, thus representing states that gradually move from erroneous states towards the standard trajectory.

[0097] Taking the continuous selection method as an example, for a time series consisting of N time states (s0, s1, …, s…), N The correction trajectory formed by the sampling and generation of training samples can be sequentially extracted using a sliding window of fixed length L. For the starting position i of each window, L consecutive states [s] within the window are taken. i , …, s i+L-1 The input to the training samples is the input state set, which includes recent historical states and the current error state; simultaneously, the K consecutive states immediately following this window are taken. i+L , …, s i+L+K-1 The target output of the training sample is the target state set, which represents the short-term ideal state evolution that the system should achieve.

[0098] The generation of the training samples described above mainly describes the method of obtaining state information in the input and output information. In addition to state information, the input part of the training samples also needs to include the observation images and task instructions corresponding to the sampling time points of the training samples. The specific acquisition method is the same as that of the existing ones, so it will not be repeated here.

[0099] The training samples generated from all the correction trajectories are referred to as the first training data. This training data originates from the correction trajectories and can therefore effectively train the correction capability. The first training data is then mixed with a large-scale standard trajectory dataset generated by the robot performing its normal tasks to form the complete training dataset for the VLA model.

[0100] Step 304: Train and generate a VLA model based on the complete training dataset.

[0101] A general VLA model is selected as the base model for training. This model can receive inputs such as images, robot states, and natural language task instructions, and output predicted states after processing.

[0102] The general VLA model is trained using the complete training dataset obtained in step 303 above. The specific training method can follow the standard procedures of various existing VLA models, such as supervised fine-tuning (SFT), without changing the model architecture or loss function, offering strong compatibility and easy extensibility. The only difference is that the input training sample set during model training includes not only training samples obtained from standard trajectories but also training samples generated based on the correction trajectory in this application. Thus, a VLA model with trajectory correction capabilities can be generated.

[0103] After training, the generated VLA model is called the "correction VLA model". Next, the "correction VLA model" is deployed to the robot system to achieve trajectory correction. Specifically, this may include the following processing:

[0104] Step 305, System Deployment and Initialization.

[0105] The trained "trajectories correction VLA model" with trajectory correction capabilities is deployed into the control loop of the actual robot system. The system continuously receives real-time sensor data (images, motion status) and task commands.

[0106] Step 306: The model receives real-time sensor data and task instructions, processes them, and generates subsequent actions for the robot to achieve trajectory correction.

[0107] The system continuously receives real-time sensor data (images, motion status) and task commands, which are then processed by the deployed "VLA model for trajectory correction." Specifically, when the input data consists of sensor data and task commands generated after an execution error, the target trajectory produced by the "VLA model for trajectory correction" tends towards the standard trajectory, effectively achieving trajectory correction.

[0108] At this point, Figure 3 The training method and trajectory correction method flowchart shown in the specific embodiment of this application concludes here. This specific embodiment, through a complete and operable technical process from data acquisition, processing, model training to online deployment, enables embodied intelligent models to acquire the ability to learn from errors and autonomously recover. This method utilizes purely error-correction behavior data for targeted training, allowing the model to learn the mapping relationship of "how to recover from an erroneous state," ultimately achieving real-time, intelligent online trajectory correction.

[0109] More specifically, the training and trajectory correction methods in the above embodiments support the selection between "in-situ correction" and "back-off correction" based on task characteristics, enabling the robot to efficiently recover from errors and replan its path, thereby significantly enhancing the system's robustness and adaptability. Simultaneously, the trained "correction VLA model" focuses on learning the "correction trajectory," which starts from the error state and ends at the standard trajectory. This design effectively avoids imitating the error execution process, ensuring the reliability of behavioral decisions. Furthermore, the correction capability training method can be directly integrated into existing training processes without modifying the VLA training framework, exhibiting good compatibility.

[0110] The above describes the specific implementation of the trajectory correction method and the correction capability training method in this application. This application also provides a correction capability training device for a robot and an embodied intelligent model, which can be used to implement the aforementioned trajectory correction method and correction capability training method. Figure 6 A schematic diagram of the basic structure of the robot provided in this application. Figure 6 As shown, the robot includes: a perception module, a decision-making and control module, and an execution module.

[0111] The perception module is used to acquire perception data in real time and send the acquired perception data and natural language commands to the decision and control module.

[0112] The decision and control module is used to input the acquired perception data and natural language commands into a pre-trained VLA model with trajectory correction capabilities after the robot makes an execution error. The VLA model processes the data and generates a sequence of corrective actions. The VLA model is obtained by learning the correction trajectory.

[0113] The execution module is used to drive the robot's end effector to sequentially execute multiple correction actions in the correction action sequence, so that the robot's actual trajectory gradually approaches and eventually returns to the standard trajectory.

[0114] Figure 7 A schematic diagram of the basic structure of the error correction training device provided in this application is shown below. Figure 7 As shown, the device includes: a correction demonstration data acquisition unit, a correction trajectory interception unit, a training data generation unit, and a training unit.

[0115] The correction demonstration data acquisition unit is used to acquire correction demonstration data. The correction demonstration data includes trajectory data of the robot performing correction actions after an execution error occurs until it returns to the standard trajectory.

[0116] The correction trajectory extraction unit is used to extract the correction trajectory from the correction demonstration data. The starting state of the correction trajectory begins with the execution error, and the state sequence contained in the correction trajectory is used to characterize the complete correction process that enables the robot to return from the error state to the standard trajectory.

[0117] The training data generation unit is used to generate first training data for training the correction capability based on the correction trajectory, and to add the first training data to the training dataset of the VLA model.

[0118] The training unit is used to train and generate a VLA model with trajectory correction capabilities based on the training dataset.

[0119] Optionally, in the trajectory correction interception unit, the method for intercepting the trajectory correction may specifically include:

[0120] Determine the starting point of the correction action in the correction demonstration data;

[0121] Extract the trajectory segment from the starting point of the correction action until the robot's action sequence coincides with the corresponding action sequence in the standard trajectory or reaches the preset endpoint as the correction trajectory.

[0122] Optionally, in the training data generation unit, the processing of generating first training data for training the correction capability based on the correction trajectory may specifically include:

[0123] Training samples are obtained from the correction trajectory and used to form the first training data.

[0124] In this context, at least one input state of the training sample contains first state information in the correction trajectory used to characterize the erroneous state that needs to be corrected.

[0125] Optionally, the correction demonstration data includes in-situ correction demonstration data and / or backtracking correction demonstration data;

[0126] Among them, the on-site correction demonstration data is the demonstration data of the robot planning a smooth trajectory to return to the standard trajectory and the next key point based on the current error state after an execution error occurs; the key point is the position point that must be reached in the standard trajectory.

[0127] The backtracking and correction demonstration data is a demonstration of how the robot, after an execution error occurs, returns to the state before the execution error occurred on the standard trajectory, starting from the execution of the correction action.

[0128] Optionally, in the error correction demonstration data acquisition unit, if the state deviation between the initial error state after the erroneous operation ends and the next key point on the standard trajectory is less than or equal to a preset threshold, then an on-site error correction strategy is determined to be adopted; if the state deviation is greater than the preset threshold, then a backtracking error correction strategy is determined to be adopted. Based on the determined error correction strategy, corresponding error correction demonstration data is collected.

[0129] Optionally, in the error correction demonstration data acquisition unit, before determining the relationship between the state deviation and the preset threshold, it is first determined whether the initial error state can be directly returned to the next key point of the standard trajectory; if the initial error state can be directly returned to the next key point of the standard trajectory, the relationship between the state deviation and the preset threshold is determined, and the error correction strategy is determined; if the initial error state cannot be directly returned to the next key point of the standard trajectory, the backtracking error correction strategy is determined.

[0130] This application also provides a computer-readable storage medium that stores instructions, which, when executed by a processor, can perform the steps in the robot trajectory correction method described above. In practical applications, the computer-readable medium may be included in the devices / apparatus / systems of the above embodiments, or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium stores instructions, which, when executed by a processor, can perform the steps in the robot trajectory correction method described above.

[0131] According to the embodiments disclosed in this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof, but not intended to limit the scope of protection of this application. In the embodiments disclosed in this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0132] Figure 7 The training device shown can be located in Figure 6 The modules shown can be either modules within the robot or standalone devices.

[0133] Figure 8 This is a schematic diagram of the basic structure of the electronic device provided in this application. Figure 8 As shown, specifically:

[0134] The electronic device may include a processor 801 with one or more processing cores, a memory 802 with one or more computer-readable storage media, and a computer program stored in the memory and executable on the processor. When the program in the memory 802 is executed, a robot trajectory correction method can be implemented.

[0135] Specifically, in practical applications, this electronic device may also include components such as a power supply 803 and an input / output unit 804. Those skilled in the art will understand that... Figure 8 The structure of the electronic device shown does not constitute a limitation on the electronic device and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. Wherein:

[0136] The processor 801 is the control center of the electronic device. It connects various parts of the electronic device through various interfaces and lines. By running or executing software programs and / or modules stored in the memory 802, and calling data stored in the memory 802, it performs various functions of the server and processes data, thereby monitoring the electronic device as a whole.

[0137] Memory 802 can be used to store software programs and modules, i.e., the aforementioned computer-readable storage medium. Processor 801 executes various functional applications and data processing by running the software programs and modules stored in memory 802. Memory 802 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created according to the use of the server, etc. In addition, memory 802 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, memory 802 may also include a memory controller to provide processor 801 with access to memory 802.

[0138] The electronic device also includes a power supply 803 that supplies power to the various components. This power supply can be logically connected to the processor 801 via a power management system, enabling functions such as charging, discharging, and power consumption management. The power supply 803 may also include one or more DC or AC power supplies, a recharging system, a power fault detection circuit, a power converter or inverter, a power status indicator, or any other components.

[0139] The electronic device may also include an input / output unit 804, which can be used to receive input digital or character information, and to generate keyboard, mouse, joystick, and optical signal inputs related to user settings and function control. The input / output unit 804 can also be used to display information input by the user or information provided to the user, as well as various graphical user interfaces, which can be composed of graphics, text, icons, video, and any combination thereof.

[0140] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A trajectory correction method for a robot, characterized in that, include: Acquire perception data and task instructions after the robot encounters an execution error; The perceived data and task instructions are input into a pre-trained VLA model with trajectory correction capabilities; wherein the VLA model is obtained by learning the correction trajectory. The VLA model is used to process the perceived data and task instructions to generate a sequence of corrective actions; The robot is controlled to execute multiple correction actions in the correction action sequence in sequence, so that the robot's actual trajectory gradually approaches and eventually returns to the standard trajectory.

2. The method according to claim 1, characterized in that, The training process of the VLA model includes: Acquire correction demonstration data; the correction demonstration data includes trajectory data of the robot starting to perform correction actions after an execution error occurs until it returns to the standard trajectory; The correction trajectory is extracted from the correction demonstration data; the starting state of the correction trajectory is the starting state for executing the correction action, and the state sequence contained in the correction trajectory is used to characterize the complete correction process that enables the robot to return from the error state to the standard trajectory. First training data for training the correction capability is generated based on the correction trajectory, and the first training data is added to the training dataset of the VLA model. A VLA model with trajectory correction capabilities is generated based on the training dataset.

3. The method according to claim 2, characterized in that, The interception and correction trajectory includes: Determine the starting point of the correction action in the correction demonstration data; The trajectory segment from the starting point of the correction action until the robot's action sequence coincides with the corresponding action sequence in the standard trajectory or reaches the preset endpoint is extracted as the correction trajectory.

4. The method according to claim 2, characterized in that, The generation of first training data for training the correction capability based on the correction trajectory includes: Training samples are obtained from the correction trajectory to form the first training data; Wherein, at least one input state of the training sample contains first state information in the correction trajectory used to characterize the error state that needs to be corrected.

5. The method according to claim 2, characterized in that, The correction demonstration data includes on-site correction demonstration data and / or backtracking correction demonstration data; The on-site correction demonstration data refers to the demonstration data of the robot planning a smooth trajectory to return to the standard trajectory and the next key point based on the current error state after an execution error occurs; the key point is the position point that must be reached in the standard trajectory. The backtracking and correction demonstration data is a demonstration of how the robot, after an execution error occurs, retreats to the position on the standard trajectory before the execution error occurred, starting from the execution of the correction action.

6. The method according to claim 5, characterized in that, Before acquiring the correction demonstration data, the training process further includes: if the state deviation between the initial error state after the error operation ends and the next key point on the standard trajectory is less than or equal to a preset threshold, then it is determined that an on-site correction strategy is adopted; if the state deviation is greater than the preset threshold, then it is determined that a backtracking correction strategy is adopted. The acquisition of correction demonstration data includes: collecting corresponding correction demonstration data based on a determined correction strategy.

7. The method according to claim 6, characterized in that, Before determining the relationship between the state deviation and the preset threshold, the training process further includes: If the initial error state can directly regress to the next key point of the standard trajectory, then the relationship between the state deviation and the preset threshold is determined, and an on-site correction strategy is adopted; if the initial error state cannot directly regress to the next key point of the standard trajectory, then a backtracking correction strategy is adopted.

8. A robot, characterized in that, include: The module consists of a perception module, a decision-making and control module, and an execution module. The perception module is used to acquire perception data in real time and send the acquired perception data and natural language commands to the decision and control module. The decision and control module is used to input the acquired perception data and natural language instructions into a pre-trained VLA model with trajectory correction capabilities after the robot makes an execution error. The VLA model then processes the data to generate a sequence of corrective actions. The VLA model is obtained by learning the correction trajectory. The execution module is used to drive the robot's end effector to sequentially execute multiple correction actions in the correction action sequence, so that the robot's actual trajectory gradually approaches and eventually returns to the standard trajectory.

9. A training device for corrective capabilities of an embodied intelligent model, characterized in that, include: The system includes a data acquisition unit for correction demonstration, a trajectory capture unit for correction, a training data generation unit, and a training unit. The correction demonstration data acquisition unit is used to acquire correction demonstration data; wherein, the correction demonstration data includes trajectory data of the robot performing correction actions after an execution error occurs until it returns to the standard trajectory; The correction trajectory extraction unit is used to extract the correction trajectory from the correction demonstration data; wherein the starting state of the correction trajectory begins with the execution error, and the state sequence contained in the correction trajectory is used to characterize the complete correction process that enables the robot to return from the error state to the standard trajectory. The training data generation unit is used to generate first training data for training the correction capability based on the correction trajectory, and to add the first training data to the training dataset of the VLA model. The training unit is used to train and generate a VLA model with trajectory correction capability based on the training dataset.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the instruction is executed by the processor, it can realize the trajectory correction method of the robot according to any one of claims 1 to 6.

11. An electronic device, characterized in that, The electronic device includes at least a computer-readable storage medium and a processor; The processor is configured to read the executable instructions from the computer-readable storage medium and execute the instructions to implement the trajectory correction method for the robot according to any one of claims 1 to 6.

Images