Autonomous operation control method for underwater robot, electronic device, and storage medium

Abstract

This application provides an autonomous operation control method, electronic device, and storage medium for an underwater robot. The autonomous operation control method includes: acquiring operating condition data collected by multiple sensors; performing feature extraction and multi-source data fusion on the operating condition data to generate a structured operating condition representation; inputting the structured operating condition representation into a lightweight language model to generate a task descriptor including task type, environmental state, and constraints; using the task descriptor to search a task experience database and match multiple candidate action sequences; performing safety verification on the multiple candidate action sequences and selecting automatic control commands that ensure the stable operation of the underwater robot; and updating the task experience database based on the execution results of the automatic control commands. The autonomous operation control method, electronic device, and storage medium of this application can improve the intelligence level and safety of underwater operations.

Description

Technical Field

[0001] This application relates to the field of underwater robot technology, and in particular to underwater robot autonomous operation control methods, electronic devices, and storage media. Background Technology

[0002] Underwater robots (ROVs / AUVs) have significant applications in marine resource exploration, underwater pipeline inspection, underwater archaeology, and military reconnaissance. However, the unstructured underwater environment, characterized by uneven lighting, low visibility, and strong current disturbances, poses a great challenge to the autonomous perception and control of these robots.

[0003] Current underwater robot control systems primarily rely on manual remote control, requiring operators to maintain high concentration for extended periods, resulting in high labor intensity and a high risk of operational errors in complex environments. While some systems have introduced rule-based assisted control, the rule base struggles to cover the diversity of unstructured environments, leading to poor adaptability.

[0004] In recent years, the development of embodied intelligence technology has provided new ideas for the autonomous control of underwater robots. Semantic reasoning capabilities based on large language models can realize the understanding and planning of complex tasks, but existing solutions have the following shortcomings: First, the temporal dimension misalignment problem exists in multi-source heterogeneous sensor data, affecting the perception accuracy; second, there is a lack of effective physical safety constraint mechanisms, which poses a risk of dynamic instability; and third, there is a lack of online experience accumulation and self-evolution capabilities, making it unable to adapt to the dynamically changing underwater environment.

[0005] Therefore, there is an urgent need for an underwater robot assisted driving system that can integrate multimodal semantic intent reasoning, physical security constraints, and online experience evolution to improve the intelligence and safety of underwater operations. Summary of the Invention

[0006] In view of the shortcomings of the prior art described above, the purpose of this application is to provide an autonomous operation control method, electronic device and storage medium for underwater robots, so as to solve the problems of high labor intensity and easy operation error in the prior art when manually operating underwater robots.

[0007] In a first aspect, this application provides an autonomous operation control method for an underwater robot, comprising: The system acquires operating condition data collected by multiple sensors, including underwater environment image data, the position and attitude data of the underwater robot, and hydrological data. Feature extraction and multi-source data fusion are performed on the aforementioned operating condition data to generate a structured operating condition characterization; The structured process condition representation is input into a lightweight language model to generate a task descriptor that includes task type, environmental state, and constraints. The task descriptor is used to search the task experience database and match multiple candidate action sequences; The multiple candidate action sequences are subjected to security verification to select the automatic control commands that ensure the stable operation of the underwater robot. The task experience database is updated based on the execution results of the automatic control instructions.

[0008] In one implementation of the first aspect, before acquiring the operating condition data collected by the multi-source sensors, the method further includes: Acquire real-time video stream data from the vision sensor; Acquire attitude data collected in real time by the attitude sensor; Visual change detection is performed on the video stream data to obtain environmental visual deviation values; Motion parameters are calculated from the attitude data to obtain the angular velocity and net acceleration of the underwater robot. The net acceleration is used to perform self-motion compensation on the environmental visual deviation value to eliminate visual interference caused by the underwater robot’s own pose change and obtain the real environmental visual deviation value. The instantaneous wake-up score is calculated by weighting and fusing the real-world visual deviation value with the angular velocity. The instantaneous wake-up score is accumulated and integrated over a preset time period to obtain the wake-up score value; When the wake-up integral value is greater than or equal to a preset trigger threshold, a joint wake-up command for waking up other sensors is generated.

[0009] In one implementation of the first aspect, the step of fusing multi-source data to generate a structured working condition characterization from the operating condition data includes: Align the time steps of operating condition data collected by sensors with different sampling frequencies to obtain aligned image data, attitude data, and hydrological data. Extract environmental image features from the aligned image data; The kinematic parameters of the underwater robot are calculated from the aligned attitude data; The environmental image features, kinematic parameters, hydrological data, and current timestamp are structured and encapsulated to obtain the structured chemical condition characterization.

[0010] In one implementation of the first aspect, the step of inputting the structured task condition representation into a lightweight language model to generate a task descriptor including task type, environment state, and constraints includes: The environmental image features, kinematic parameters, and hydrological data in the structural chemical condition characterization are converted into text form to obtain the state description text. The status description text is concatenated according to a preset prompt word template to generate structured prompt text; The structured prompt text and preset task prompt words are input into the lightweight language model to generate the task descriptor.

[0011] In one implementation of the first aspect, the step of using the task descriptor to retrieve a task experience database and match multiple candidate action sequences includes: Generate a task feature vector based on the task descriptor; Calculate the cosine similarity between the task feature vector and each historical experience vector in the task experience database, and filter out multiple historical experience vectors with a cosine similarity greater than or equal to a first threshold to form a heuristic reference set. Extract the historical success rate corresponding to each historical experience vector from the heuristic reference set; Extract the three-dimensional spatial coordinates of the target point and the attitude quaternion of the underwater robot at the target point from the environmental state of the task descriptor; Extract the maximum allowed speed and maximum execution time of the task from the constraints of the task descriptor; The three-dimensional spatial coordinates, the attitude quaternion, the maximum motion speed, and the maximum execution time are concatenated to obtain the matching feature vector; Extract skill records corresponding to the heuristic reference set from the preset skill pool to obtain a candidate skill set; Extract the estimated energy consumption corresponding to each skill record from the candidate skill set; Calculate the Mahalanobis distance between the matching feature vector and the feature statistics parameters of each skill record in the candidate skill set to obtain a distance metric. The comprehensive score for each skill record is calculated based on the distance metric, the historical success rate, the estimated energy consumption, and the current battery percentage of the underwater robot. The action sequences corresponding to multiple skill records whose comprehensive scores are greater than or equal to the second threshold are used as the candidate action sequences.

[0012] In one implementation of the first aspect, the step of performing security checks on the plurality of candidate action sequences to filter out automatic control commands that ensure the stable operation of the underwater robot includes: The desired thrust of each thruster of the underwater robot is generated based on the candidate action sequence; The desired thrust is combined to obtain the initial auxiliary thrust vector; Obtain the steady-state sustaining thrust vector output by the controller of the underwater robot; The initial auxiliary thrust vector and the steady-state sustaining thrust vector are vector-superimposed to obtain the superimposed total thrust vector. Calculate the infinite norm of the superimposed total thrust vector; The infinity norm is compared with a preset safety constraint threshold. When the infinity norm is greater than the safety constraint threshold, the elements of the initial auxiliary thrust vector are reduced proportionally until the infinity norm of the superimposed total thrust vector calculated based on the reduced initial auxiliary thrust vector is less than or equal to the safety constraint threshold, thus obtaining the target auxiliary thrust vector. When the infinity norm is less than or equal to the safety constraint threshold, the initial auxiliary thrust vector is directly determined as the target auxiliary thrust vector; The target auxiliary thrust vector is converted into the automatic control command.

[0013] In one implementation of the first aspect, after selecting the automatic control commands that ensure the stable operation of the underwater robot, a command arbitration step is further included: Obtain manual control signals; Determine the operational intensity value of the manual control signal; When the operation intensity value is greater than the preset intensity threshold, a manual control command is generated based on the manual operation signal, and the manual control command is used to override the automatic control command, and the underwater robot enters the manual operation mode. When the operation intensity value is less than or equal to the preset intensity threshold, and the duration of the manual control signal exceeds the preset recovery time, the automatic control command overrides the manual control command, and the underwater robot enters the automatic driving mode.

[0014] In one implementation of the first aspect, updating the task experience database based on the execution result of the automatic control instruction includes: By comparing the operating data before and after executing the automatic control command, the operating deviation is obtained; Calculate the task achievement rate based on the aforementioned working condition deviations; Compare whether the task achievement rate is greater than or equal to a preset quality threshold; If so, a completion status feature vector is generated based on the working condition data after the automatic control command is executed, and the skill identifier, task achievement degree, and completion status feature vector corresponding to the automatic control command are structurally encapsulated in a preset data format to generate experience to be injected. Calculate the similarity between the experience to be injected and the existing experience in the task experience database. When the similarity is less than a preset deduplication threshold, store the experience to be injected into the task experience database. The task completion rate is used as an immediate reward, and the action value function weights corresponding to the skill identifiers in the task experience database are iteratively updated based on a reinforcement learning algorithm.

[0015] Secondly, this application provides an electronic device, including a memory and a processor, wherein the processor is configured to execute a computer program stored in the memory to cause the electronic device to perform the steps of the underwater robot autonomous operation control method.

[0016] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the underwater robot autonomous operation control method described above.

[0017] As described above, the underwater robot autonomous operation control method, electronic device, and storage medium described in this application have at least one of the following beneficial effects: (1) A lightweight large language model is used to realize deep semantic understanding and active reasoning of unstructured underwater environment, thereby improving the ability to autonomously complete complex tasks. The lightweight language model is a 7B parameter scale model quantized by INT8. Edge reasoning is implemented on Raspberry Pi 5B through the llama.cpp framework, and the reasoning delay does not exceed 500 milliseconds.

[0018] (2) The thrust infinite norm constraint is adopted as a physical hard constraint safety fallback mechanism to ensure that the assisted driving command does not destroy the underlying dynamic steady state, thus fundamentally guaranteeing the safety of the system; (3) By strengthening the learning reward mechanism, the system can achieve online self-evolution and experience accumulation, enabling it to continuously adapt to the dynamically changing underwater environment; (4) The joint wake-up mechanism is adopted to realize asynchronous on-demand data acquisition, which significantly reduces the energy consumption of edge computing and extends the underwater operation endurance. (5) The linear interpolation time alignment method is adopted to effectively eliminate the problem of misalignment of time series dimensions of multi-source heterogeneous sensors and improve the sensing accuracy. Attached Figure Description

[0019] Figure 1 The flowchart shown is a process for an autonomous operation control method for an underwater robot as described in one embodiment of this application.

[0020] Figure 2 The flowchart shown is a joint wake-up preprocessing procedure as described in one embodiment of this application.

[0021] Figure 3 The flowchart shown is a process for generating a structured chemical condition characterization by fusing multi-source data as described in one embodiment of this application.

[0022] Figure 4The flowchart shown is a process of inputting a structural condition representation into a lightweight language model to generate a task descriptor in one embodiment of this application.

[0023] Figure 5 The flowchart shown is a process for retrieving candidate action sequences from a task experience database using a task descriptor, according to an embodiment of this application.

[0024] Figure 6 The flowchart shown is an automatic control instruction for security verification and screening of candidate action sequences in one embodiment of this application.

[0025] Figure 7 This is a flowchart illustrating the instruction arbitration steps described in one embodiment of this application.

[0026] Figure 8 This is a flowchart illustrating the updating of the task experience database based on the execution result of automatic control instructions, as described in one embodiment of this application.

[0027] Figure 9 The diagram shown is a structural schematic of an electronic device described in one embodiment of this application. Detailed Implementation

[0028] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0029] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0030] Please refer to Figure 1 In one embodiment, this application provides an autonomous operation control method for an underwater robot, comprising the following steps: Step S100: Acquire working condition data collected by multi-source sensors. The working condition data includes underwater environment image data, underwater robot position data and attitude data, and hydrological data. Step S200: Feature extraction and multi-source data fusion are performed on the operating condition data to generate a structured operating condition characterization. Step S300: Input the structured task condition representation into the lightweight language model to generate a task descriptor that includes task type, environmental state, and constraints. Step S400: Use the task descriptor to retrieve the task experience database and match multiple candidate action sequences; Step S500: Perform safety verification on multiple candidate action sequences and filter out automatic control commands that ensure the stable operation of the underwater robot; Step S600: Update the task experience database based on the execution results of the automatic control instructions.

[0031] In step S100, the multi-source sensors specifically refer to the depth camera, monocular camera, nine-axis attitude sensor, depth sensor, and pH sensor deployed on the underwater robot body. The operational data acquisition frequency is asynchronously distributed. The structured operational condition characterization specifically refers to a standard JSON object encapsulated after time-based alignment and feature encoding of the multi-source heterogeneous data. This object includes an aligned base timestamp, a 256-dimensional visual global pooling vector, DRACO compressed point cloud, quaternion attitude, water depth, water quality indicators, and semantic tag fields. The lightweight language model specifically refers to a large-scale 7B language model using INT8 quantization compression, deployed on the edge control unit, and run through an inference framework, with inference latency strictly controlled within 500 milliseconds. The task experience database specifically refers to an indexed edge-side vector repository. Each experience is embedded using 64-dimensional PCA dimensionality reduction, with a maximum capacity of 512 entries, used for rapid retrieval of historical successful cases. The candidate action sequence specifically refers to a 12-dimensional action tensor matching from a preset skill pool, containing the robotic arm joint angle sequence and eight-pusher thrust timing instructions. The automatic control command specifically refers to the PWM signal command that is mapped to a pulse width modulation of 1100 to 1900 microseconds after safety constraint calculation.

[0032] By constructing a collaborative architecture of asynchronous perception and edge semantic reasoning, this embodiment overcomes the limitations of traditional underwater control systems that rely on fixed rule bases, enabling robots to achieve high-level semantic understanding and action intent generation in unstructured, low-visibility environments. The combination of multimodal data fusion and lightweight models effectively reduces the computational load on edge computing units while ensuring millisecond-level response speeds, providing a highly real-time cognitive decision-making foundation for closed-loop autonomous operations.

[0033] Please refer to Figure 2 In one embodiment, a joint wake-up preprocessing procedure is included before acquiring the operating condition data collected by multiple sensors: Step S210: Acquire video stream data collected in real time by the visual sensor; Step S220: Obtain attitude data collected in real time by the attitude sensor; Step S230: Perform visual change detection on the video stream data to obtain the environmental visual deviation value; Step S240: Perform motion parameter calculation on the attitude data to obtain the angular velocity and net acceleration of the underwater robot; Step S250: Use net acceleration to perform self-motion compensation on the environmental visual deviation value, eliminate visual interference caused by the underwater robot's own pose change, and obtain the real environmental visual deviation value. Step S260: The visual deviation value of the real environment and the angular velocity are weighted and fused to calculate the instantaneous wake-up score; Step S270: Accumulate and integrate the instantaneous wake-up score within a preset time period to obtain the wake-up score value; Step S280: When the wake-up integral value is greater than or equal to the preset trigger threshold, a joint wake-up command for waking up other sensors is generated.

[0034] In steps S230 to S260, the environmental visual deviation value refers to the squared Euclidean distance between the current frame's visual feature vector and the background feature vector. Motion compensation is achieved through inverse dynamics. The net acceleration is obtained by calculating the underlying steady-state thrust through the pseudo-inverse of the thrust distribution matrix and subtracting the gravity component of the body coordinate system. Then, combined with the camera intrinsic parameter matrix, the visual Jacobian matrix, and a typical exposure time of 8 milliseconds, second-order compensation is applied to the pixel displacement to eliminate motion blur artifacts caused by the underwater robot's own acceleration and deceleration.

[0035] Instantaneous wake-up score is calculated using the joint wake-up function: in, Represents the wake-up function value. This represents the visual change weighting factor, with a value of 1.0. This represents the angular velocity weighting factor, with a value of 0.5. This represents the compensated visual feature vector of the current frame. Represents the background visual feature vector. Represents the current angular velocity vector. This represents the maximum angular velocity threshold, with a value of 2.0 radians per second. The preset trigger threshold is set to 2.5; when the integral of the sliding window exceeds this value, synchronous acquisition across all sensors is triggered.

[0036] This mechanism employs an on-demand activation strategy, maintaining a low-power standby state during periods of environmental stability and waking up the high-frequency acquisition link only during sudden visual changes or drastic attitude fluctuations. The self-motion compensation module effectively eliminates the interference of the body's maneuvers on visual perception, ensuring that wake-up decisions only respond to changes in the actual external environment. This significantly reduces the overhead of invalid data processing and extends the endurance of underwater equipment during a single operation.

[0037] Please refer to Figure 3 In one embodiment, multi-source data fusion is performed on the operating condition data to generate a structured operating condition characterization, including: Step S310: Align the time steps of the operating condition data collected by sensors with different sampling frequencies to obtain aligned image data, attitude data and hydrological data. Step S320: Extract environmental image features from the aligned image data; Step S330: Calculate the kinematic parameters of the underwater robot from the aligned attitude data; Step S340: The environmental image features, kinematic parameters, hydrological data and the current timestamp are structured and encapsulated to obtain a structured chemical condition characterization.

[0038] In step S310, time alignment is performed at 100Hz for the nine-axis attitude sensor.

[0039] The timestamp is a global reference clock. For visual frames with a timing deviation of less than or equal to 10 seconds, a rigid body motion compensation rotation matrix is ​​first constructed using angular velocity. Considering the line-by-line readout characteristics of the depth camera, the full frame readout time is 33 milliseconds. The actual exposure time offset is calculated row by row, and the coordinates of the compensated feature points are projected. Finally, the features of each row are uniformly mapped to the reference time. The alignment interpolation formula is as follows: in, The interpolated estimate represents the value at the target time. and These represent the original sensor measurements at adjacent sampling times. This represents the alignment reference moment. Kinematic parameters include the three-dimensional coordinates of the position and the unit quaternion. The structured encapsulation process serializes the fields of each sub-object according to a predefined pattern and writes them to an asynchronous circular buffer.

[0040] This alignment process, through hardware-level timing deviation modeling and software-level linear interpolation compensation, completely eliminates the timing dimension misalignment problem caused by differences in sampling rates and exposure mechanisms in multi-source heterogeneous sensors. The snapshot generation mechanism with a unified time reference ensures the consistency and spatial synchronization of subsequent semantic inference and motion control input data, laying a data foundation for high-precision pose calculation.

[0041] Please refer to Figure 4 In one embodiment, the structured task condition representation is input into a lightweight language model to generate a task descriptor including task type, environmental state, and constraints, including: Step S410: Convert the environmental image features, kinematic parameters and hydrological data in the structural chemical condition characterization into text form to obtain the state description text; Step S420: Concatenate the status description text according to the preset prompt word template to generate structured prompt text; Step S430: Input the structured prompt text and preset task prompt words into the lightweight language model to generate a task descriptor.

[0042] In steps S410 and S420, the preset prompt word template specifically refers to serializing visual semantic labels, point cloud centroid coordinates, position and pose quaternions, water depth, battery percentage, and pH value into a prompt text that mixes natural language and numerical values ​​in a fixed field order. In step S430, the model output uses a grammatical constraint decoding mechanism. During the autoregressive generation process, the sampler dynamically filters out illegal words based on the current grammatical state. in, This represents the probability of generating a lexical unit at step t given the historical context and grammatical state. Represents the original logarithmic probability vector. Represents the syntax mask function. The mechanism sets the log probability of disallowed terms in the current state to negative infinity, which forces the model output to strictly conform to the JSON format of the target, state, and constraint triplet.

[0043] By employing hard-constraint decoding through grammatical sampling, this embodiment eliminates the format drift and hallucination output problems commonly encountered when deploying large language models at the edge. A structured cue word concatenation strategy explicitly embeds high-dimensional physical states into the semantic space, enabling the model to understand the spatial mapping relationship between the underwater robot's pose and environmental elements. The 100% legality guarantee of the output format eliminates the need for additional fault-tolerant parsing in the downstream skill matching module, ensuring that end-to-end inference latency remains stable within a 500-millisecond threshold, meeting the requirements of real-time control timing.

[0044] Please refer to Figure 5 In one embodiment, a task experience database is retrieved using a task descriptor to match multiple candidate action sequences, including: Step S510: Generate a task feature vector based on the task descriptor; Step S520: Calculate the cosine similarity between the task feature vector and each historical experience vector in the task experience database, and filter out multiple historical experience vectors with a cosine similarity greater than or equal to the first threshold to form a heuristic reference set. Step S530: Extract the historical success rate corresponding to each historical experience vector from the heuristic reference set; Step S540: Extract the three-dimensional spatial coordinates of the target point and the attitude quaternion of the underwater robot at the target point from the environmental state of the task descriptor. Step S550: Extract the maximum allowed speed and maximum execution time of the task from the constraints of the task descriptor; Step S560: Concatenate the three-dimensional spatial coordinates, attitude quaternion, maximum motion speed, and maximum execution time to obtain the matching feature vector; Step S570: Extract skill records corresponding to the heuristic reference set from the preset skill pool to obtain a candidate skill set; Step S580: Extract the estimated energy consumption corresponding to each skill record from the candidate skill set; Step S590: Calculate the Mahalanobis distance between the matching feature vector and the feature statistics parameters of each skill record in the candidate skill set to obtain the distance metric. Step S591: Calculate the comprehensive score of each skill record based on the distance metric, historical success rate, estimated energy consumption, and the current battery percentage of the underwater robot. Step S592: Select the action sequences corresponding to multiple skill records with a comprehensive score greater than or equal to the second threshold as candidate action sequences.

[0045] In step S520, the first threshold is set to 0.85. Cosine similarity calculation is based on a 64-dimensional PCA-reduced embedding vector. In step S570, the skill pool is stored using a structured vector library, where each record contains an identifier, skill name, action sequence, step interval, depth constraint, speed limit, estimated energy consumption, and historical feature mean vector. Inverse covariance matrix. Mahalanobis distance matching uses the inverse covariance matrix to adaptively weight the sensitivity of spatial pose to constraints. The overall score is calculated as follows: in, The overall evaluation score representing skill records, Represents the Mahalanobis distance metric. Represents historical success rate. This represents the battery degradation weighting factor, with a value of 0.3. Represents estimated energy consumption. This represents the upper limit benchmark for energy consumption, with a value of 500 joules. This represents the low battery penalty factor, with a value of 0.4. This represents the current battery percentage.

[0046] The second threshold is set as the baseline for comprehensive scoring. When the experience base is full, a min-heap strategy based on task achievement is used to eliminate low-value experiences, and new experiences are directly discarded if their cosine similarity with the nearest neighbor in the base is greater than 0.99.

[0047] This two-stage retrieval mechanism decouples semantic generalization for initial screening from physical constraint-based fine ranking. Cosine similarity quickly converges to semantically similar historical trajectories, while Mahalanobis distance combined with covariance structure accurately assesses the physical match between the current working condition and skill applicability. Dynamically introduced power weights and energy consumption penalty factors automatically favor low-power, high-success-rate skill sequences under low-power conditions. Deduplication and min-heap elimination strategies ensure the high quality and lightweight nature of the experience base, avoiding the drag on retrieval performance from redundant information.

[0048] Please refer to Figure 6 In one embodiment, a safety check is performed on multiple candidate action sequences to filter out automatic control commands that ensure the stable operation of the underwater robot, including: Step S610: Generate the desired thrust of each thruster of the underwater robot based on the candidate action sequence; Step S620: Combine the desired thrusts to obtain the initial auxiliary thrust vector; Step S630: Obtain the steady-state sustaining thrust vector output by the underwater robot's controller; Step S640: The initial auxiliary thrust vector and the steady-state maintenance thrust vector are vector superimposed to obtain the superimposed total thrust vector; Step S650: Calculate the infinite norm of the superimposed total thrust vector; Step S660: Compare the infinite norm with the preset safety constraint threshold; Step S670: When the infinity norm is greater than the safety constraint threshold, the elements of the initial auxiliary thrust vector are reduced proportionally until the infinity norm of the superimposed total thrust vector calculated based on the reduced initial auxiliary thrust vector is less than or equal to the safety constraint threshold, thus obtaining the target auxiliary thrust vector. Step S680: When the infinite norm is less than or equal to the safety constraint threshold, the initial auxiliary thrust vector is directly determined as the target auxiliary thrust vector. Step S690: Convert the target auxiliary thrust vector into an automatic control command.

[0049] In step S630, the steady-state sustaining thrust vector is calculated in real time by the PD self-balancing controller operated by the edge master control unit: in, Represents the underlying steady-state thrust vector. and Represents the diagonal gain matrix. This represents the six-dimensional pose mixing error vector.

[0050] The safety constraint threshold setting follows the thrust infinity norm constraint: in, Represents the safety constraint threshold. This represents the safety margin factor, with a value of 0.85. This represents the rated maximum thrust of the thruster, with a forward thrust of 19.6 Newtons.

[0051] In step S690, the 8 thrust values ​​are converted into PWM pulse widths after pseudo-inverse mapping through the thrust allocation matrix. The forward thrust mapping formula is: The reverse thrust mapping formula is: The data frame is encapsulated as a header, length, floating-point array and checksum, and sent to the underlying controller at a specified communication cycle.

[0052] This dynamic safety envelope mechanism strictly limits the auxiliary thrust within the physical capacity and dynamic steady-state boundaries of the thruster. When the superposition of auxiliary commands and underlying self-balancing commands may cause thrust saturation or attitude divergence, the linear attenuation module automatically scales proportionally to ensure that control commands never disrupt the underwater robot's six-degree-of-freedom dynamic balance. The hard constraint fallback design eliminates the risk of underwater equipment instability caused by planned actions exceeding limits, providing physical-level safety assurance for autonomous operation.

[0053] Please refer to Figure 7 In one embodiment, after selecting the automatic control commands that ensure the stable operation of the underwater robot, a command arbitration step is also included: Step S710: Obtain manual control signal; Step S720: Determine the operation intensity value of the manual control signal; Step S730: When the operation intensity value is greater than the preset intensity threshold, a manual control command is generated based on the manual control signal, and the manual control command overrides the automatic control command, and the underwater robot enters the manual operation mode. In step S740, when the operation intensity value is less than or equal to the preset intensity threshold and the duration of the manual control signal exceeds the preset recovery time, the automatic control command overrides the manual control command, and the underwater robot enters the autonomous driving mode.

[0054] In step S720, the operation intensity value specifically refers to the Euclidean norm amplitude of the 6-dimensional target speed or position command input by the operator via the remote control terminal. The preset intensity threshold is set to 0.1. When the amplitude exceeds the dead zone, the control loop immediately deprives the automatic command authority and switches to manual takeover; when the command falls back into the dead zone and the duration exceeds 2 seconds, the system determines that the operator's intervention intention has ended and smoothly returns control to the automatic driving loop.

[0055] The dead-zone detection and timeout recovery logic establishes a top-priority arbitration mechanism for human-machine collaboration. In emergency situations such as sudden water flow disturbances, approaching obstacles, or equipment malfunctions, the operator can instantly take over the equipment with a small joystick movement. The system's zero-delay response overrides automatic commands, ensuring absolute reliability of emergency intervention. Simultaneously, the preset recovery time and dead-zone anti-shake design prevent high-frequency oscillations in control between automatic and manual modes, achieving a seamless and smooth transition of human-machine authority.

[0056] Please refer to Figure 8 In one embodiment, updating the task experience database based on the execution result of the automatic control command includes: Step S810: Compare the operating condition data before and after executing the automatic control command to obtain the operating condition deviation; Step S820: Calculate the task achievement rate based on the working condition deviation; Step S830: Compare whether the task achievement rate is greater than or equal to the preset high-quality threshold. Step S840, if yes, then generate a completion status feature vector based on the working condition data after the automatic control command is executed, and encapsulate the skill identifier, task achievement degree and completion status feature vector corresponding to the automatic control command in a structured manner according to a preset data format to generate the experience to be injected; Step S850: Calculate the similarity between the experience to be injected and the existing experience in the task experience database. When the similarity is less than the preset deduplication threshold, store the experience to be injected into the task experience database. Step S860: The task completion rate is used as an immediate reward, and the weights of the action value function corresponding to the skill identifiers in the task experience database are iteratively updated based on the reinforcement learning algorithm.

[0057] In step S820, the task achievement score calculation integrates the three-dimensional residuals of position, attitude, and target state: in, Represents the degree of task completion. The function represents the interval truncated to zero and one. These represent the weights for position, attitude, and target state deviations, with values ​​of 0.4, 0.3, and 0.3, respectively. Represents positional residuals, in meters, based on a normalized datum. It is 0.5 meters. Represents the quaternion attitude deflection angle, in radians, with a normalized reference. for , This represents the deviation of the target state. If the capture is successful, it is 0 or 1. If the detection task is successful, it is 1 minus the intersection and union ratio.

[0058] The default high-quality threshold is set to 0.8.

[0059] In step S860, reinforcement learning employs a linear function approximation combined with the Q-learning algorithm. The state space is 270-dimensional, which is then projected to a 64-dimensional feature space via offline PCA. in, The value of an action performed in a given state. This represents a 64-dimensional weight vector maintained independently by skill index. This represents the feature vector after dimensionality reduction. The online iterative update formula is as follows: in, This represents the learning rate, with a value of 0.01. Represents an instant reward. This represents the discount factor, with a value of 0.95. This represents the state vector after execution.

[0060] This evaluation and iterative closed-loop system achieves rigorous filtering of experience quality and online evolution of policy weights. Only high-quality trajectories with an achievement level exceeding the quality threshold are injected into the memory, while noisy experiences are discarded. Linear approximation learning enables low-computational-overhead error backpropagation on resource-constrained edge devices, dynamically adjusting skill scheduling probabilities during continuous operations to gradually approach the globally optimal control policy. Without offline retraining, the robot can continuously adapt to changes in water flow, equipment wear, and environmental feature drift as the number of task executions increases, possessing online self-evolution capabilities.

[0061] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For apparatus embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments. Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions.

[0062] The scope of protection for the underwater robot autonomous operation control method described in this application is not limited to the execution order of the steps listed in this embodiment. Any solution implemented by adding, subtracting, or replacing steps in the prior art based on the principles of this application is included within the scope of protection of this application.

[0063] Please refer to Figure 9 This embodiment also provides an electronic device, which is a user's mobile device such as a mobile phone, PAD, wearable device, or smart AI device; the electronic device includes a memory for storing computer programs; and a processor for running the computer programs to implement the underwater robot autonomous operation control method in the above embodiments.

[0064] The memory is connected to the processor via a system bus and they communicate with each other. The memory is used to store computer programs, and the processor is used to run the computer programs so that the electronic device executes the underwater robot autonomous operation control method of the above embodiments.

[0065] It should also be noted that the system bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This system bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. The communication interface is used to enable communication between the database access device and other devices (such as clients, read-write databases, and read-only databases).

[0066] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0067] In addition, this embodiment also provides a storage medium storing program instructions, which, when executed by a processor, implement the underwater robot autonomous operation control method described in the above embodiment.

[0068] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing a processor. The program can be stored in a computer-readable storage medium, which is a non-transitory medium, such as random access memory, read-only memory, flash memory, hard disk, solid-state drive, magnetic tape, floppy disk, optical disk, and any combination thereof. The storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. This available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video disc (DVD)), or a semiconductor medium (e.g., solid-state drive (SSD)).

[0069] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0070] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for autonomous operation control of an underwater robot, characterized in that, include: The system acquires operating condition data collected by multiple sensors, including underwater environment image data, the position and attitude data of the underwater robot, and hydrological data. Feature extraction and multi-source data fusion are performed on the aforementioned operating condition data to generate a structured operating condition characterization; The structured process condition representation is input into a lightweight language model to generate a task descriptor that includes task type, environmental state, and constraints. The task descriptor is used to search the task experience database and match multiple candidate action sequences; The multiple candidate action sequences are subjected to security verification to select the automatic control commands that ensure the stable operation of the underwater robot. The task experience database is updated based on the execution results of the automatic control instructions.

2. The underwater robot autonomous operation control method according to claim 1, characterized in that, Before acquiring the operating condition data collected by the multi-source sensors, the following is also included: Acquire real-time video stream data from the vision sensor; Acquire attitude data collected in real time by the attitude sensor; Visual change detection is performed on the video stream data to obtain environmental visual deviation values; Motion parameters are calculated from the attitude data to obtain the angular velocity and net acceleration of the underwater robot. The net acceleration is used to perform self-motion compensation on the environmental visual deviation value to eliminate visual interference caused by the underwater robot’s own pose change and obtain the real environmental visual deviation value. The instantaneous wake-up score is calculated by weighting and fusing the real-world visual deviation value with the angular velocity. The instantaneous wake-up score is accumulated and integrated over a preset time period to obtain the wake-up score value; When the wake-up integral value is greater than or equal to a preset trigger threshold, a joint wake-up command for waking up other sensors is generated.

3. The underwater robot autonomous operation control method according to claim 2, characterized in that, The step of fusing multi-source data to generate a structured working condition characterization includes: Align the time steps of operating condition data collected by sensors with different sampling frequencies to obtain aligned image data, attitude data, and hydrological data. Extract environmental image features from the aligned image data; The kinematic parameters of the underwater robot are calculated from the aligned attitude data; The environmental image features, kinematic parameters, hydrological data, and current timestamp are structured and encapsulated to obtain the structured chemical condition characterization.

4. The underwater robot autonomous operation control method according to claim 3, characterized in that, The step of inputting the structured task condition representation into a lightweight language model to generate a task descriptor including task type, environment state, and constraints includes: The environmental image features, kinematic parameters, and hydrological data in the structural chemical condition characterization are converted into text form to obtain the state description text. The status description text is concatenated according to a preset prompt word template to generate structured prompt text; The structured prompt text and preset task prompt words are input into the lightweight language model to generate the task descriptor.

5. The underwater robot autonomous operation control method according to claim 1, characterized in that, The step of using the task descriptor to retrieve the task experience database and match multiple candidate action sequences includes: Generate a task feature vector based on the task descriptor; Calculate the cosine similarity between the task feature vector and each historical experience vector in the task experience database, and filter out multiple historical experience vectors with a cosine similarity greater than or equal to a first threshold to form a heuristic reference set. Extract the historical success rate corresponding to each historical experience vector from the heuristic reference set; Extract the three-dimensional spatial coordinates of the target point and the attitude quaternion of the underwater robot at the target point from the environmental state of the task descriptor; Extract the maximum allowed speed and maximum execution time of the task from the constraints of the task descriptor; The three-dimensional spatial coordinates, the attitude quaternion, the maximum motion speed, and the maximum execution time are concatenated to obtain the matching feature vector; Extract skill records corresponding to the heuristic reference set from the preset skill pool to obtain a candidate skill set; Extract the estimated energy consumption corresponding to each skill record from the candidate skill set; Calculate the Mahalanobis distance between the matching feature vector and the feature statistics parameters of each skill record in the candidate skill set to obtain a distance metric. The comprehensive score for each skill record is calculated based on the distance metric, the historical success rate, the estimated energy consumption, and the current battery percentage of the underwater robot. The action sequences corresponding to multiple skill records whose comprehensive scores are greater than or equal to the second threshold are used as the candidate action sequences.

6. The underwater robot autonomous operation control method according to claim 1, characterized in that, The step of performing security checks on the multiple candidate action sequences and filtering out the automatic control commands that ensure the stable operation of the underwater robot includes: The desired thrust of each thruster of the underwater robot is generated based on the candidate action sequence; The desired thrust is combined to obtain the initial auxiliary thrust vector; Obtain the steady-state sustaining thrust vector output by the controller of the underwater robot; The initial auxiliary thrust vector and the steady-state sustaining thrust vector are vector-superimposed to obtain the superimposed total thrust vector. Calculate the infinite norm of the superimposed total thrust vector; The infinity norm is compared with a preset safety constraint threshold. When the infinity norm is greater than the safety constraint threshold, the elements of the initial auxiliary thrust vector are reduced proportionally until the infinity norm of the superimposed total thrust vector calculated based on the reduced initial auxiliary thrust vector is less than or equal to the safety constraint threshold, thus obtaining the target auxiliary thrust vector. When the infinity norm is less than or equal to the safety constraint threshold, the initial auxiliary thrust vector is directly determined as the target auxiliary thrust vector; The target auxiliary thrust vector is converted into the automatic control command.

7. The underwater robot autonomous operation control method according to claim 1, characterized in that, After selecting the automatic control commands that ensure the stable operation of the underwater robot, a command arbitration step is also included: Obtain manual control signals; Determine the operational intensity value of the manual control signal; When the operation intensity value is greater than the preset intensity threshold, a manual control command is generated based on the manual operation signal, and the manual control command is used to override the automatic control command, and the underwater robot enters the manual operation mode. When the operation intensity value is less than or equal to the preset intensity threshold, and the duration of the manual control signal exceeds the preset recovery time, the automatic control command overrides the manual control command, and the underwater robot enters the automatic driving mode.

8. The underwater robot autonomous operation control method according to claim 1, characterized in that, Updating the task experience database based on the execution result of the automatic control command includes: By comparing the operating data before and after executing the automatic control command, the operating deviation is obtained; Calculate the task achievement rate based on the aforementioned working condition deviations; Compare whether the task achievement rate is greater than or equal to a preset quality threshold; If so, a completion status feature vector is generated based on the working condition data after the automatic control command is executed, and the skill identifier, task achievement degree, and completion status feature vector corresponding to the automatic control command are structurally encapsulated in a preset data format to generate experience to be injected. Calculate the similarity between the experience to be injected and the existing experience in the task experience database. When the similarity is less than a preset deduplication threshold, store the experience to be injected into the task experience database. The task completion rate is used as an immediate reward, and the action value function weights corresponding to the skill identifiers in the task experience database are iteratively updated based on a reinforcement learning algorithm.

9. An electronic device comprising a memory and a processor, characterized in that, The processor is used to execute the computer program stored in the memory to cause the electronic device to perform the steps of the underwater robot autonomous operation control method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the autonomous operation control method for an underwater robot as described in any one of claims 1 to 8.

Images