Control method, device and equipment for scanning robot driven by interpretation intention and medium

Abstract

The invention relates to the technical field of medical image intelligent processing and robot control, and discloses an interpretation intention-driven scanning robot control method, device, equipment and medium, and the method comprises the steps: obtaining a scanning image, motion state data and task interaction reference data, generating a visual feature set and an interpretation intention vector, and performing time sequence coding on the motion state data to obtain historical track representation, generating motion decision features based on cross-modal feature fusion, obtaining a high-level strategy instruction and a low-level control instruction through hierarchical decoding, and synthesizing a driving signal for controlling the scanning robot to execute scanning operation. According to the method, cross-modal fusion is carried out on the visual interpretation intention and the motion trail features, so that the robot can make an action decision based on the operation intention, blind scanning behaviors are reduced, the objectivity and stability of the scanning process are improved, and automatic scanning control more conforming to clinical interpretation logic is achieved.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image processing and robot control technology, and in particular to a scanning robot control method, device, equipment and medium driven by interpretation intent. Background Technology

[0002] Ultrasound scanning robots in clinical assisted imaging tasks typically rely on rule-based search strategies or end-to-end imitation learning models for automatic path planning and posture adjustment. However, these methods often learn only from the appearance of the image or historical movement patterns, failing to construct interpretation logic that matches actual clinical procedures and ignoring the cognitive sequence of "understanding the target first, then deciding how to move" in human operation. The lack of clear intent expression makes the robot prone to ambiguous actions when performing scanning tasks, such as blind movement in situations with blurred tissue structures, poor acoustic windows, or unclear boundaries, affecting scanning efficiency and imaging stability.

[0003] Some studies have attempted to introduce reinforcement learning frameworks, but these rely on visual rewards or external localization systems, often making them highly sensitive to fluctuations in image quality and insufficiently adaptable to changes in task scenarios. Training requires a large number of interactive samples, resulting in low learning efficiency and difficulty in meeting the real-time and reliability requirements of practical scanning tasks. Furthermore, these methods typically cannot explain the decision-making basis behind their actions, making it difficult to provide clinically understandable behavioral logic in scenarios involving lesion search, complex anatomical structure recognition, or path adjustment.

[0004] Existing methods generally use images as the sole input for strategy, failing to incorporate implicit intentions reflected in the operator's target judgments, window selection preferences, or interactive behaviors during the scanning process. This single-condition driven strategy ignores meso-level decision-making information closely related to action selection in clinical practice, resulting in a lack of stable, clear, and generalizable action basis for robots when faced with anatomical differences, image noise variations, or local abnormal structures. This limits the effectiveness of automated ultrasound systems in complex scenarios. Summary of the Invention

[0005] The main objective of this invention is to provide a scanning robot control method, device, equipment, and storage medium driven by interpretation intent. This invention aims to solve the technical problem that existing technologies cannot explicitly combine the operator's intent and the robot's motion state during the scanning process to construct a decision chain of interpretation before operation, resulting in blind scanning and ambiguous actions due to a lack of clear intent basis for scanning actions.

[0006] To achieve the above objectives, the present invention provides a scanning robot control method driven by intent interpretation, comprising: Real-time acquisition of scan image sequences, end effector motion status data, and task interaction reference data generated during the scanning process; The scanned image sequence is convolved using a visual feature extraction network to generate a set of visual features containing multi-resolution information. The task interaction reference data is mapped to a conditional query vector, and the conditional query vector is used as an attention guidance signal to perform attention inference calculation on the visual feature set to generate a judgment intent vector that represents the judgment decision logic. Extract sequence segments of the motion state data within a preset time window, and encode the sequence segments with temporal features to generate a historical trajectory representation containing motion inertial information; Using the interpretation intent vector as a decoding condition, a cross-modal feature fusion operation is performed on the historical trajectory representation to generate action decision features; The action decision features are input into the hierarchical decoding module for parallel prediction to generate high-level policy instructions containing discrete policy categories and low-level control instructions containing continuous motion parameters. The high-level strategy instructions and the low-level control instructions are combined to form a drive signal, which controls the scanning robot to perform the corresponding scanning operation.

[0007] Furthermore, to achieve the above objectives, the present invention provides a scanning robot control device driven by intent interpretation, comprising: The data acquisition module is used to acquire in real time the scan image sequence, motion status data of the end effector, and task interaction reference data generated during the scanning process; The visual feature extraction module is used to perform convolution processing on the scanned image sequence using a visual feature extraction network to generate a set of visual features containing multi-resolution information. The intent reasoning module is used to map the task interaction reference data into a condition query vector, and use the condition query vector as an attention guidance signal to perform attention reasoning calculation on the visual feature set to generate a judgment intent vector that represents the judgment decision logic. The trajectory encoding module is used to extract the sequence segments of the motion state data within a preset time window, and to encode the sequence segments with temporal features to generate a historical trajectory representation containing motion inertial information. The feature fusion module is used to perform cross-modal feature fusion operation on the historical trajectory representation using the interpretation intent vector as a decoding condition to generate action decision features; The hierarchical decoding module is used to input the action decision features into the hierarchical decoding module for parallel prediction, and generate high-level policy instructions containing discrete policy categories and low-level control instructions containing continuous motion parameters. The execution control module is used to synthesize a drive signal based on the high-level strategy instructions and the low-level control instructions, and control the scanning robot to perform corresponding scanning operations.

[0008] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a scanning robot control program driven by intent interpretation stored in the memory and executable on the processor, wherein when the scanning robot control program driven by intent interpretation is executed by the processor, it implements the steps of the scanning robot control method driven by intent interpretation as described above.

[0009] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a scanning robot control program driven by intent interpretation, wherein the scanning robot control program driven by intent interpretation, when executed by a processor, implements the steps of the scanning robot control method driven by intent interpretation as described above.

[0010] Beneficial Effects: This invention relates to the field of intelligent medical image processing and robot control technology, and discloses a scanning robot control method, device, equipment, and medium driven by interpretation intent. The method includes: real-time acquisition of scanned image sequences, end effector motion state data, and task interaction reference data; generation of a multi-resolution visual feature set using a visual feature extraction network; mapping the task interaction reference data into conditional query vectors and performing attention inference calculations to generate interpretation intent vectors; temporal feature encoding of motion state data to obtain historical trajectory representations; cross-modal feature fusion using the interpretation intent vector as decoding conditions to generate action decision features; parallel generation of high-level strategy instructions and low-level control instructions through a hierarchical decoding module; and synthesis of drive signals based on these two types of instructions to control the scanning robot to perform scanning operations. This invention, by fusing visual interpretation intent with motion trajectory features across modalities, enables the robot to make action decisions based on operational intent, thereby reducing blind scanning and improving the targeting and stability of the scanning process, achieving automated scanning control that is more in line with clinical interpretation logic. Attached Figure Description

[0011] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for an intent-driven scanning robot control method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the intention-driven scanning robot control method of the present invention; Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the intention-driven scanning robot control device of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0012] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0013] The intent-driven scanning robot control method provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can obtain scanned image sequences, end effector motion state data, and task interaction reference data in real time from the client. It uses a visual feature extraction network to generate a multi-resolution visual feature set, maps the task interaction reference data into conditional query vectors, performs attention inference calculations to generate interpretation intent vectors, performs temporal feature encoding on the motion state data to obtain historical trajectory representations, and uses the interpretation intent vector as a decoding condition to complete cross-modal feature fusion to generate action decision features. A hierarchical decoding module generates high-level policy instructions and low-level control instructions in parallel, and synthesizes drive signals based on these two types of instructions to control the scanning robot to perform scanning operations. This invention, by fusing visual interpretation intent and motion trajectory features across modalities, enables the robot to make action decisions based on operational intent, thereby reducing blind scanning and improving the targeting and stability of the scanning process, achieving more automated scanning control that conforms to clinical interpretation logic. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0014] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the intent-driven scanning robot control method provided by the present invention. It should be noted that although the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0015] like Figure 2 As shown, the intention-driven scanning robot control method proposed in this invention includes the following steps: S10, real-time acquisition of scan image sequences, motion state data of the end effector, and task interaction reference data generated during the scanning process; In this embodiment, the scanned image sequence originates from image frames continuously output by the scanning imaging device. These image frames are typically generated at a fixed frequency and carry acquisition time information. After establishing a data path through the video acquisition interface, the image frames are cached as a continuous sequence in chronological order, ensuring the continuity of imaging changes over time. Motion state data includes the position, posture, and mechanical information of the end effector, which is fed back in real time by the robot control system. Position is often represented by three-dimensional coordinates, posture can be expressed using Euler angles or quaternions, and mechanical information is acquired by force sensors installed on the probe or end effector structure to reflect contact changes on the scanning surface. Task interaction reference data reflects the operator's intention information. It captures the screen gaze point through an eye-tracking device, records interactive actions through the input interface, and combines these with behavioral markers to form a unified data structure that can be used to express scanning intentions. The above three types of data record the acquisition time with their respective timestamps and are mapped to the same time base through resampling or interpolation strategies, so that visual, motion, and interaction information form a synchronized input set.

[0016] This embodiment acquires image sequences, motion state information, and interactive reference information simultaneously. The input data can simultaneously reflect imaging changes, probe posture dynamics, and operator intent, enabling subsequent processing to have a complete spatiotemporal awareness foundation and improving the stability and interpretability of the scanning task.

[0017] S20, the scanned image sequence is convolved using a visual feature extraction network to generate a set of visual features containing multi-resolution information; In this embodiment, the visual feature extraction network is a deep structure constructed by stacking convolutional computation units. Each frame in the scanned image sequence first undergoes input preprocessing, mapping the pixel matrix into a data format suitable for network computation. During convolutional processing, convolutional layers of different depths within the network perform region scanning operations with different receptive fields and strides. This makes shallow convolutional layers more sensitive to fine-grained information such as local brightness changes and edge contours, while deep convolutional layers, after multiple convolutional stacks, form an abstract representation of global anatomical morphology, tissue distribution, and texture structure. Due to the differences in spatial size, number of channels, and representational content among the convolutional layers, the network naturally forms a multi-resolution representational structure during the overall computation process.

[0018] Since scanned image sequences change continuously over time, batch normalization, activation functions, and spatial scale adjustment operations are incorporated into the convolution process to obtain stable visual representations. Layer-by-layer convolution aggregates feature outputs at different resolutions into a single set, preserving both the overall layout and local details of anatomical structures. This set includes both feature maps compressed from the original resolution through convolution and abstract representations generated by deep convolutions. Each feature map contains different semantic responses along the channel dimension, reflecting acoustic window morphology, tissue boundary clarity, and local texture variations.

[0019] To avoid information loss due to scale shifts during convolution, multi-level sampling paths can be set in the network, allowing the image to participate in computation in convolutional layers at different scales, thus maintaining the consistency of visual representation across different spatial scales. The generated visual feature set consists of multiple resolution levels, and there is no mandatory fixed structure between the levels; instead, they are automatically learned by the network during training, enabling the system to utilize visual information at different scales for subsequent judgments.

[0020] This embodiment generates a multi-resolution visual feature set through convolution processing, enabling the network to simultaneously represent local tissue changes and overall anatomical structures in the image. This avoids insufficient information caused by relying solely on single-scale images and improves the robustness and interpretation stability of the system under complex ultrasound imaging conditions.

[0021] S30, the task interaction reference data is mapped into a conditional query vector, and the conditional query vector is used as an attention guidance signal to perform attention inference calculation on the visual feature set to generate a judgment intent vector representing the judgment decision logic; In this embodiment, the task interaction reference data includes the operator's behavioral information during the imaging and interpretation process, such as screen gaze point position, input commands, and stages of the diagnostic process. This information reflects the target area the operator is focusing on at each moment, the meaning of the operation they wish to trigger, and the working context they are in. Therefore, these inputs need to be converted into vector representations that can participate in computation. The mapping process is completed jointly by spatial coding units, semantic embedding units, and stage embedding units, enabling the original behavioral signals to express attention focus, semantic information, and process progress in vector space.

[0022] The conditional query vector is formed by concatenating the aforementioned multi-source embedding vectors. Its function is to provide the network with intent cues such as "what the operator is paying attention to," "what the operator is going to do," and "which stage they are currently in," acting as a query term in attention inference. The visual feature set provides key-value representations from the image content, and the conditional query vector determines the attention weight by calculating similarity with these image features, thereby guiding the image information with behavioral information.

[0023] Attentional inference computation is implemented through a multi-head attention structure, where multiple attention heads focus on visual responses at different spatial scales or semantic dimensions, enabling the system to simultaneously capture various types of visual cues such as salient regions, anatomical boundaries, and texture variations. The weighted results generated by each attention head are aggregated to form an intermediate interpretation vector. To improve consistency between the representation and the task stage, a gating adjustment mechanism is introduced, converting the stage embedding vector into adjustment weights and applying them to the intermediate interpretation vector, making the final output more consistent with the current working context. For example, the global structure is strengthened in the "localization stage," while local details are emphasized in the "detailed scanning stage."

[0024] Once the interpretation intent vector is calculated, it becomes a set of abstract expressions that can be used to reflect the operator's judgment logic. It contains three types of information: spatial attention bias, action preference, and process context, enabling subsequent models to understand the meaning of the interpretation rather than relying solely on superficial changes in the image.

[0025] This embodiment transforms task interaction reference data into conditional query vectors and participates in attention inference, enabling behavioral information to directly guide the aggregation direction of visual features. This transforms the interpretation model from relying solely on image content to making judgments based on the operator's intent, thereby reducing misfocusing in ambiguous areas, improving the ability to identify key areas, and enhancing the system's adaptability and stability at different scanning stages.

[0026] S40, extract the sequence segment of the motion state data within a preset time window, and encode the sequence segment with temporal features to generate a historical trajectory representation containing motion inertial information; In this embodiment, motion state data is a type of numerical sequence that is continuously updated over time, typically including spatial position, orientation, contact force, or other parameters reflecting the motion state of the end effector. This data is inherently time-dependent, therefore data segments need to be constructed within continuous time periods to recover the dynamic characteristics of the motion process. A preset time window serves as the boundary of the time range, limiting the length of each segment to ensure that the resulting sequence segments cover sufficient motion trajectories and maintain temporal continuity. This window can be set based on the sampling frequency, robot response cycle, or motion change characteristics.

[0027] Sequence segments are multidimensional sequences obtained by cropping motion state data according to time windows. The timestamps within these segments maintain an increasing relationship, reflecting temporal information such as translational, angular, and force changes during the motion process. Segments typically contain multiple channel dimensions, therefore, dimension alignment and numerical standardization are needed to maintain channel scale comparability, enabling subsequent encoding processes to extract shared representational patterns from multi-source motion parameters.

[0028] Temporal feature encoding uses learnable or rule-based temporal processing networks to compress and represent sequence fragments, mapping high-dimensional, long-sequence motion data into shorter vector forms. The encoding process captures short-term dynamic changes and long-term trends through continuous convolution, recursive updates, temporal operations, or the superposition of positional markers, enabling the final generated vector to express the directionality, persistence, and rate of change of motion inertia. This vector, known as the historical trajectory representation, typically contains abstract expressions of the continuity of motion direction, velocity change patterns, force variation laws, and the smoothness of pose transitions.

[0029] Inertial information here refers to the correlation characteristics between consecutive actions during motion, such as whether the displacement direction remains consistent, whether the angle change shows an acceleration trend, and whether the force feedback exhibits periodic changes. Through the encoding process, the local change patterns in the sequence are concentrated into a unified vector, thus providing the system with an abstract description of the robot's motion trend over a period of time.

[0030] This embodiment extracts motion inertia information from continuous motion by extracting continuous motion state data into a sequence within a time window and performing time-series encoding. This enables the system to identify motion trends, rates of change, and force feedback patterns, providing a trajectory expression with temporal continuity for subsequent motion reasoning. This reduces interference from random fluctuations and enhances the ability to depict real motion patterns.

[0031] S50, using the interpretation intent vector as a decoding condition, perform a cross-modal feature fusion operation on the historical trajectory representation to generate action decision features; In this embodiment, the processing content uses the interpretation intent vector as the decoding condition. This vector is used as a guiding signal to apply to the historical trajectory representation to construct a fusion expression that reflects the current interpretation requirements. The interpretation intent vector originates from the preceding analysis stage and is the result of encoding information such as scanning task requirements, interactive input preferences, or target areas of interest. It is typically composed of a set of real numbers and can be represented as a fixed-length vector structure, the dimensionality of which is determined by the preceding extraction network. The historical trajectory representation is the temporal encoding result of the motion state data. It is a compressed expression of the scanning robot's motion trends, speed changes, direction adjustments, and inertial characteristics. It is also represented as a vector array with a certain dimension, used to reflect the motion behavior over a period of time.

[0032] To fuse two datasets from different sources within the same representation space, the interpretation intent vector needs to be adapted in terms of numerical structure and scale, enabling it to influence the representation weights of the trajectory representation as a guiding condition. During fusion, methods such as vector alignment, dimension mapping, additive adjustment, or dot product adjustment can be used to allow the interpretation intent vector to influence the numerical changes of trajectory features at the feature level, forming a directional guidance consistent with task requirements. This process not only aligns the shapes of the two types of vectors but also implicitly encodes their semantic influence relationships, giving the trajectory representation the ability to respond to interpretation intent after fusion. For example, when certain dimensions in the interpretation intent vector represent a preference for steady motion, the fusion process will strengthen the numerical regions in the trajectory representation corresponding to steady trends, thus making the fusion result more aligned with interpretation requirements.

[0033] Cross-modal feature fusion refers to the process of converting data streams from different modalities (i.e., interpretation intent and motion trajectory) into a unified, jointly processable representation. Here, "cross-modal" emphasizes the differences between the two types of data in terms of source structure, temporal attributes, and semantic abstraction level, combining them in the feature space through a fusion mechanism. The resulting action decision feature is a comprehensive vector describing the robot's motion tendency under the current interpretation requirement. This feature space can simultaneously express motion inertia and task preference, giving it decision-oriented characteristics and providing a foundation for subsequent action generation.

[0034] This embodiment utilizes the interpretation intent vector to guide the trajectory representation fusion process, enabling the generated action decision features to possess additional interpretation preference information on top of the motion trend expression. This helps improve the consistency between the robot's motion trajectory and task requirements, reduces redundant adjustments caused by random actions, and makes the subsequent control process more directional and interpretation-related.

[0035] S60, the action decision features are input into the hierarchical decoding module for parallel prediction to generate high-level policy instructions containing discrete policy categories and low-level control instructions containing continuous motion parameters. In this embodiment, the action decision features, as an expression structure derived from the fusion processing, contain abstract information about historical trajectory trends and response information to interpretation preferences, typically represented as a set of transformed numerical vectors. When generating control decisions, this vector needs to be input into a hierarchical decoding module to extract discrete category information that guides the overall scanning behavior and continuous parameter information that quantifies the specific execution action from different decision levels. The hierarchical decoding module is a structure that differentiates different output types based on the same input. It contains decoding paths designed for different prediction targets; higher-level paths output policy categories, and lower-level paths generate continuous control quantities. Discrete policy categories refer to a type of labeled output with decision semantics, such as representing propulsion, attitude adjustment, position maintenance, or movement in a specific direction. Essentially, it is a selection from a finite set; therefore, the action decision features need to be mapped to a probability distribution space to complete the category determination. Continuous motion parameters encode the velocity, attitude change, offset angle, or other continuous values ​​required for the actual execution action. They depend on the numerical gradient in the action decision features to generate quantized control signals that can be directly used by the actuator.

[0036] After inputting action decision features, the hierarchical decoding module needs to adjust their numerical structure to match the parameter scale and computational requirements of each decoding path. Decoding paths used to generate discrete policy categories typically include a mapping layer and a distribution generation layer, which select policy categories by mapping the vector structure to a class space. Decoding paths used to generate continuous motion parameters generate target parameters through a continuous numerical regression structure to reflect the changing trends of robot actions in intensity, amplitude, and direction. Parallel prediction emphasizes that the two decoding paths are structurally independent but share the same input, ensuring that the two types of outputs maintain semantic complementarity while avoiding mutual interference, thereby achieving coordination between high-level planning and low-level control at the feature level.

[0037] This embodiment generates high-level strategy categories and low-level control parameters based on the same action decision features, enabling the scanning robot to form a clear division of labor between global direction selection and local position adjustment, improving the structured nature of decision output, reducing coupling interference caused by single path prediction, and making the overall action control more stable and more relevant to interpretation.

[0038] S70: Based on the high-level strategy instructions and the low-level control instructions, a drive signal is synthesized to control the scanning robot to perform the corresponding scanning operation.

[0039] In this embodiment, high-level strategy instructions are a type of motion planning information given in categorical form, typically consisting of a finite set. They are used to express the scanning robot's directional decisions regarding overall orientation, operational targets, or motion trends, such as indicating movement direction, posture adjustment intention, or field-of-view expansion intention. Low-level control instructions are a type of execution quantity information represented in continuous numerical form, used to describe the specific amplitude, speed, offset, or posture adjustment of motion as continuous parameters. They originate from a detailed mapping of the numerical structure in the motion decision characteristics. The process of synthesizing high-level strategy instructions and low-level control instructions into a drive signal requires structural coordination between the two types of instructions. This ensures that categorical decisions have a regulatory role in the final drive, while numerical control quantities can cover actual execution requirements. This synthesis process typically requires establishing an instruction mapping relationship, converting categorical decisions into influence weights or motion templates, so that they can work together with low-level continuous parameters on the control interface. The drive signal is an instruction format that the motor controller or actuator can directly parse. It contains the target values ​​required for direction control, speed control, position adjustment, or posture compensation. Through unified encoding of the high-level and low-level instruction structures, independent control quantities that the actuator can recognize can be formed. The process of the robot performing scanning operations relies on the continuous transmission of these drive signals, which enable the actuators to move, rotate, or adjust pressure according to the expected trajectory, thereby adapting to the human anatomy and obtaining a stable scanning path.

[0040] By establishing a classification instruction mapping table, high-level policy instructions can be transformed into action templates, and low-level control variables can be modified according to the parameter structure of these templates to maintain consistency in directionality. For example, if a high-level policy instruction corresponds to directional movement, then all values ​​of the low-level control variables only take effect in the corresponding directional dimension. Alternatively, a weight fusion method can be used to multiply the weights corresponding to the high-level policy categories with the low-level control parameters, enabling the driving signal to produce a stronger directional response when the category decision is significant. For scenarios requiring high precision in the actuator, independent control signal encoding methods can be established for different categories, giving the driving signal a differentiated structure under different task semantics and improving trajectory adjustment capabilities. During the continuous delivery phase, interpolation can be combined with the execution cycle to keep the action smooth and reduce mechanical vibration. In structurally complex anatomical regions, the adjustment capability of classification weights can be appropriately increased to generate a stronger path correction trend in areas with poor posture.

[0041] This embodiment synthesizes categorical strategy instructions and continuous control parameters at the signal level, enabling the decision-making system to simultaneously possess directional planning capabilities and detailed adjustment capabilities, thereby obtaining structured and hierarchical control outputs and improving the coherence of scanning actions and target relevance.

[0042] In one embodiment, step S10 above includes: S101, establish a communication connection with the scanning imaging device using the video acquisition interface, and acquire video stream frame images output by the scanning imaging device in real time according to a preset sampling frequency through the video acquisition interface, and arrange the video stream frame images in chronological order of acquisition time to generate a scanned image sequence. S102, connect to the robot control system through the data communication interface, and read the six-degree-of-freedom pose parameters fed back by the robot control system in real time through the data communication interface. Collect real-time contact force parameters using a six-dimensional force sensor installed on the end effector. Combine the six-degree-of-freedom pose parameters with the real-time contact force parameters to generate motion state data of the end effector. S103: The eye-tracking sensor captures the screen gaze coordinates when the operator looks at the display terminal, and the human-computer interaction interface receives the operation command signal and workflow stage identifier input by the operator. The screen gaze coordinates, the operation command signal and the workflow stage identifier are encapsulated and processed to generate task interaction reference data. S104, respectively obtain the image frame timestamp of the scanned image sequence, the state timestamp of the motion state data of the end effector, and the interaction timestamp of the task interaction reference data, and set the image frame timestamp as the reference time axis; S105, using a linear interpolation algorithm, the motion state data of the end effector and the task interaction reference data are resampled based on the state timestamp and the interaction timestamp, and the motion state data of the end effector and the task interaction reference data are mapped onto the reference time axis.

[0043] In this embodiment, real-time acquisition of the scanned image sequence first relies on establishing a stable communication connection between the video acquisition interface and the scanning imaging device. The video acquisition interface can be a network interface, a serial bus interface, or a logical interface exposed by a dedicated image acquisition card. Through this interface, continuously output video stream frames are received from the imaging device. The preset sampling frequency is given in the system parameters; it can be consistent with the imaging device's frame rate, or configured as an integer or fractional multiple. Internal buffering and frame dropping strategies ensure that the acquisition thread outputs frame images at a uniform frequency. The acquisition thread appends an acquisition time to each frame image, monotonically increments the sampling time, and stores the frame images in chronological order in a queue, thus forming a structurally continuous scanned image sequence. This sequence serves as a temporal reference for visual information in subsequent processing.

[0044] The motion state data of the end effector is obtained through a bidirectional communication channel established with the robot control system via a data communication interface. The data communication interface can be an Ethernet-based industrial communication interface, a fieldbus interface, or a serial port-based control interface. The interface protocol includes a message format that periodically returns six-DOF pose parameters. The six-DOF pose parameters typically include position and attitude components in three-dimensional space. The position component is given in three-dimensional coordinates in the robot's end effector coordinate system or world coordinate system, while the attitude component is parameterized using Euler angles, quaternions, or direction cosine matrices. A six-dimensional force sensor mounted on the end effector outputs measured values ​​of contact force and contact torque within the same communication cycle. This data is read by an acquisition thread, and unit conversion and zero-bias correction are performed according to the sensor calibration parameters. The pose parameters and contact force parameters are paired by time stamps, and motion state data at a single point in time is constructed through vector concatenation or structural combination, forming a time-varying motion state sequence that structurally includes both position and attitude information as well as contact force change information.

[0045] Task interaction reference data is generated jointly by an eye-tracking sensor and a human-computer interaction interface. The eye-tracking sensor faces the working area of ​​the display terminal, continuously sampling the operator's gaze point and outputting the gaze coordinates in the screen coordinate system through an internal algorithm. These coordinates can be represented in pixel or normalized coordinate form. The human-computer interaction interface connects to the operator's input device, receiving button input, touch operations, or command menu selections. Each valid input is parsed into structured operation command signals, such as mode switching, freezing the screen, and adjusting depth. The workflow management module maintains an identifier for the current inspection stage, indicating different stages of the inspection process, such as coarse scanning, local magnification, and fine measurement. The screen gaze coordinates, operation command signals, and workflow stage identifiers are packaged and stored simultaneously, forming task interaction reference data through a unified data structure. This ensures that each record at any given time includes three types of information: gaze focus position, interaction action, and process context.

[0046] The time alignment process aligns the image sequence, motion state sequence, and task interaction sequence using a unified time reference. In the scanned image sequence, each frame is timestamped at the time of acquisition; in the motion state sequence, each set of pose and contact force data is timestamped; and in the task interaction sequence, each interaction record is timestamped. The system selects the image frame timestamp as the reference timeline from these three types of timestamps because subsequent processing typically revolves around interpreting and making decisions based on each frame, requiring the image generation time as a unified reference. The reference timeline is treated as a set of discrete time points, and a linear interpolation algorithm resamples the motion state data and task interaction reference data at these time points. During linear interpolation, the system first finds the two adjacent sampling points in the state timestamp and interaction timestamp sequences that are adjacent to the current reference time. Then, the adjacent sampling values ​​are weighted according to the time ratio to generate the interpolation result corresponding to the reference time, thus aligning the motion state data with the image frames in time. The task interaction reference data can be resampled using nearest neighbor preservation or linear interpolation. When the interaction event is a discrete trigger event, the validity label of the most recent event can be preserved within adjacent time periods. After resampling, both motion state data and task interaction reference data are mapped onto a timeline consistent with the timestamps of the image frames, thus forming a multi-source synchronous data structure based on image frames.

[0047] This embodiment acquires and aligns scanned image sequences, end effector motion state data, and task interaction reference data synchronously on a unified timeline. This allows for the simultaneous acquisition of visual information, mechanical motion information, and human-computer interaction information at the time point corresponding to each image frame. This enables subsequent interpretation of intent modeling and motion control learning to have a precise temporal correspondence at the data level, reducing the interference of time drift between different data sources on the decision-making process and improving the ability to characterize the causal relationship between the doctor's operational behavior and image changes.

[0048] In one embodiment, step S20 above includes: S201, the image frames in the scanned image sequence are input into the backbone convolution module of the visual feature extraction network, and the backbone convolution module is used to perform a step-by-step downsampling convolution operation on the image frames to generate multiple low-level feature maps with different spatial dimensions and channel dimensions. S202, using the feature pyramid module of the visual feature extraction network, select the high-level semantic feature map in the bottom feature map for upsampling processing, and enlarge the spatial resolution of the high-level semantic feature map to match the low-level geometric feature map in the bottom feature map. S203, the channel dimension of the low-level geometric feature map is adjusted by the horizontal connection channel, and the adjusted low-level geometric feature map is superimposed with the upsampled high-level semantic feature map to generate a fused feature map that integrates semantic information and spatial details. S204, the fused feature map is smoothed by using a convolution kernel to eliminate the upsampling aliasing effect, and multiple pyramid feature maps of different scales are output. The multiple pyramid feature maps of different scales are then combined to form a visual feature set containing multi-resolution information.

[0049] In this embodiment, each frame of the scanned image sequence is sequentially input into the backbone convolutional module of the visual feature extraction network. The backbone convolutional module can be composed of multiple layers of two-dimensional convolutional units, nonlinear activation units, and spatial downsampling units stacked sequentially. Weighted accumulation operations of local regions are achieved by sliding convolutional kernels within the image plane, outputting response values ​​in the channel dimension. Spatial downsampling is typically achieved through convolution or pooling operations with a stride greater than one, reducing the feature map size by a fixed ratio in the spatial dimension, thereby aggregating a larger range of contextual information on a coarser grid. As the network depth increases, the backbone convolutional module generates multiple feature maps level by level. These feature maps gradually decrease in spatial size and gradually increase in the number of channels. Lower-level feature maps are more sensitive to geometric details such as edges, textures, and brightness changes, while higher-level feature maps aggregate more semantic information such as anatomical structures, organ outlines, and acoustic window regions. This series of feature maps with different spatial sizes and channel dimensions together constitutes the bottom-level feature map set.

[0050] After obtaining the set of low-level feature maps, a feature pyramid module is introduced to reorganize the features at different levels. The feature pyramid module selects several high-level semantic feature maps with smaller spatial dimensions but higher semantic abstraction from the low-level feature maps, and uses an upsampling operator to interpolate and restore them in the spatial dimension, thus enlarging the output features in the width and height directions to match the larger low-level geometric features. Figure 1 Upsampling operators can employ bilinear interpolation, nearest-neighbor interpolation, or transposed convolution, etc., to generate semantic features with higher spatial resolution by inserting interpolation points or performing deconvolution operations on a two-dimensional grid. Simultaneously, the low-level geometric feature map is mapped to the same channel dimension as the upsampled semantic feature map through lateral connection channels. Lateral connection channels typically use a single-layer one-to-one convolution to achieve a linear transformation in the number of channels, ensuring the additivity of the two inputs in terms of channel count. After channel adjustment, the adjusted low-level geometric feature map and the upsampled high-level semantic feature map are added or concatenated element-wise, thus preserving both fine spatial details and aggregated semantic information in a single feature map. This superposition result is denoted as the fused feature map.

[0051] To mitigate jagged edges and checkerboard artifacts generated during upsampling, smooth convolution is applied to the fused feature maps. Smoothing convolution uses a small kernel with a stride of one to locally combine responses from adjacent regions, making the responses of adjacent pixels more continuous and suppressing excessively high-frequency noise components. After smooth convolution, multiple feature maps with different spatial sizes but relatively balanced semantic representations are obtained according to different source layers. These feature maps can be arranged into a hierarchical structure, i.e., a pyramid feature map set, in order from high resolution to low resolution. By repeatedly performing high-level feature upsampling, low-level feature lateral mapping, fusion stacking, and smooth convolution in the same network, pyramid feature maps of several scales can be output, each scale corresponding to a specific spatial sampling interval and receptive field range. Finally, all pyramid feature maps at all scales are uniformly organized into a visual feature set in the data structure. The visual feature set retains the correspondence between each scale and the original image frames in the scanned image sequence, providing multi-resolution input for subsequent intent-based attention inference and cross-modal fusion.

[0052] This embodiment utilizes a backbone convolution module and a feature pyramid module to perform stepwise downsampling, cross-layer fusion, and smoothing on the scanned image sequence. This allows for the simultaneous acquisition of a multi-scale feature set covering both fine-grained texture information and large-scale structural information in a single forward propagation. This enables the alignment of subsequent interpretation intent with trajectory information to no longer be limited to a single spatial resolution. Consequently, even when lesion boundaries are blurred, acoustic windows are limited, or there are significant differences in the scale of anatomical structures, the region relevant to clinical interpretation can be captured more stably. This improves the adaptability of subsequent intent modeling and action decision-making processes to complex ultrasound images.

[0053] In one embodiment, step S30 above includes: S301, parse the task interaction reference data and extract the screen gaze point coordinates, operation command signals and workflow stage identifiers; S302, the screen gaze point coordinates are mapped to spatial position embedding vectors using a position encoder, and the operation command signal and the workflow stage identifier are mapped to semantic command embedding vectors and stage state embedding vectors respectively using a semantic embedding layer. S303, perform feature concatenation processing on the spatial location embedding vector, the semantic instruction embedding vector and the stage state embedding vector to obtain the conditional query vector; S304, using the conditional query vector as the query term and mapping the visual feature set as a key value term, perform multi-head cross-attention calculation, and weighted aggregate the visual feature set by calculating the matching degree between the query term and the key value term to obtain intermediate interpretation features; S305, using a gating network to calculate prior adjustment weights based on the stage state embedding vector, and using the prior adjustment weights to weight and correct the intermediate interpretation features to generate a interpretation intent vector.

[0054] In this embodiment, the task interaction reference data includes screen gaze coordinates, operation command signals, and workflow stage identifiers. These three types of information reflect the spatial focus position, operation semantics, and the current diagnostic process stage, respectively. When parsing the task interaction reference data, the screen coordinate sequence output by the eye-tracking module can be decoded into two-dimensional coordinate points in the display terminal coordinate system, the commands generated by the operator's input device can be encoded into discrete symbol sequences, and the stage identifiers recorded by the process control system can be mapped into discrete state markers. All three are then combined into the same data structure under a unified time index, facilitating subsequent joint encoding within the same time step.

[0055] The position encoder converts screen gaze coordinates into spatial position embedding vectors. In implementation, the two-dimensional screen coordinates are first normalized to a preset interval, and then a fixed-dimensional vector representation is generated through linear transformation or a combination of sine and cosine basis functions, ensuring that different spatial positions have distinguishable geometric relationships in the embedding space. The spatial position embedding vector preserves the operator's gaze offset trend on the image plane during subsequent calculations, such as a bias towards the intercostal region, the apex of the heart, or a specific blood vessel neighborhood. The semantic embedding layer converts operation command signals and workflow stage identifiers into continuous vector forms. Operation command signals can originate from button commands, touch operations, or voice recognition output. By querying an embedding dictionary or a small feedforward network, a semantic command embedding vector can be assigned to each operation command to represent the type of action the operator expects to perform, such as widening the acoustic window, fine-tuning the probe angle, or rescanning suspicious areas. Workflow stage identifiers can indicate different process stages, such as coarse search, local fine-tuning, measurement recording, or result confirmation. Stage state embedding vectors are generated through the same embedding mapping to distinguish the attention preferences for image regions at different stages.

[0056] The spatial location embedding vector, semantic instruction embedding vector, and stage state embedding vector are concatenated into a single high-dimensional vector during feature concatenation, aligning spatial, semantic, and procedural information along the feature dimension. In practice, the three vectors can be concatenated sequentially along the channel dimension, and the dimension of the concatenated vector can be adjusted using a linear mapping according to network structure requirements to match the input dimension of the subsequent attention calculation module. The resulting high-dimensional vector is defined as a conditional query vector at this stage, used to initiate a conditional retrieval within the visual feature set, expressing the imaging content the operator wishes to focus on at the current time, current operation, and current stage.

[0057] In attention inference computation, the conditional query vector serves as the query term, forming a query-key-value relationship with the feature units in the visual feature set. The visual feature set, already organized into feature vectors of spatial location and scale in the previous stage, is mapped to key vectors and corresponding value vectors through linear transformation, forming a key-value term set. During multi-head cross-attention computation, the conditional query vector is input to multiple attention heads, each generating a query representation in a different subspace using independent linear projection. For each attention head, the matching degree between the query representation and each key vector is calculated using dot product or other similarity metrics, yielding the degree of preference of the conditional query vector for different visual locations or scales within that subspace. The matching degree is scaled and normalized to form a set of weight coefficients. This set of weights is then weighted and summed with the corresponding value vectors to output the aggregation result for that attention head. The aggregation results from all attention heads are concatenated or summed along the channel dimension to obtain the intermediate interpretation features under the combined effect of multiple subspaces. Intermediate interpretation features numerically reflect the alignment between task interaction reference data and visual feature set, that is, which image regions and scales are most important for subsequent decisions under the current gaze, instruction and stage conditions.

[0058] Gated networks are used to introduce process-related prior conditioning after intermediate readability feature generation. In implementation, a gated network can employ several fully connected layers and nonlinear activation units, taking the stage state embedding vector as input and outputting a set of prior conditioning weights. These weights can be scalars or vectors with the same dimensions as the intermediate readability features, used to modulate the intermediate readability features across different dimensions. When the prior conditioning weights are in vector form, they can be applied to the intermediate readability features through element-wise multiplication to enhance or suppress different semantic channels. For example, in the coarse search stage, the weights for the large-scale structural channels can be increased, while in the fine search stage, the weights for the high-resolution lesion detail channels can be enhanced. The intermediate readability features, after being weighted and corrected by the prior conditioning weights, are defined as a readability intent vector. This vector contains the potential readability decision logic jointly determined by spatial fixation, operational intent, and process stage, and is output as a fixed-dimensional vector for subsequent cross-modal fusion with historical trajectory representations.

[0059] Furthermore, during the offline training phase, the intent generation module receives multimodal training samples, including ultrasound image sequences, task interaction reference data, and view category labels, structural visibility labels, lesion suspicion labels, and next target direction labels for the same time index. The conditional query vector and visual feature set are input into the cross-attention unit at the current time index, and aggregated using the attention formula: The intermediate readout sequence is obtained. The stage prior vector and the multi-source context are input into the gating network, and the gating attention is applied according to the stage prior gating attention formula: Calculate the fused attention distribution, apply the fused attention to the visual feature set, and then perform aggregation operations: The interpretation intent vector is obtained. To enable the interpretation intent vector to simultaneously constrain view recognition, structural visibility assessment, lesion suspicion estimation, and prediction of the next target direction, a multi-task loss function is introduced: By continuously updating the relevant parameters of the interpretation module through backpropagation, the same network can output a stable interpretation intent vector through forward computation only during the online inference stage after training, which is then used for subsequent action decision processing.

[0060] in, This represents the time index, corresponding to the current moment in the scanning process. The conditional query vector is a one-dimensional vector derived from task interaction reference data, which is embedded and concatenated to characterize the current diagnostic intent and area of ​​interest. This represents the feature sequence after the visual feature set is expanded at time t. Each element corresponds to a spatial location or patch on the image, containing local texture and structural information. Attn(·) represents a multi-head cross-attention operation unit, which takes a query, key, and value vector set as input and generates an attention output sequence through a weighted summation. The intermediate interpretation represents the feature sequence output by the attention unit, which has already incorporated salient region information from the conditional query and visual features. `Pool(·)` represents the aggregation operation unit, which can be average pooling, weighted pooling, or self-attention-based pooling, used to convert the sequence into a single data structure. Compress it into a fixed-length vector. This represents the intent vector to be interpreted, which will be used as a decoding condition in subsequent modules for action generation. d represents the dimension of the intent vector.

[0061] The visual context encoding for time step t is generally obtained by pooling or compressing the feature map of the current ultrasound image. It can be understood as a global vector representation of the content of the current frame image. The multimodal prior encoding for time step t, such as the prior vector obtained by fusing information from eye-tracking heatmaps, ROI regions, operation instructions, and stage identifiers, is used to reflect the operator's subjective focus area and current task intent. (·) indicates that it operates on visual context encoding. The nonlinear transformation on the network can be implemented as a one- or multi-layer fully connected network, and the output is an intermediate representation used for summation with the prior branches in the same space. (·) indicates that it acts on multimodal prior coding. Nonlinear transformations on the surface can also be achieved using fully connected networks or other feedforward structures to map prior vectors to a space with the same dimension as the visual branch. Indicates the effect on The learnable weight matrix (or weight vector) is used to compress the intermediate representation of the visual branch into scalar or element-wise weights of an attention weight map. Indicates the effect on The learnable weight matrix (or weight vector) on the prior branch is used to compress the intermediate representation of the prior branch to a size similar to the prior branch. The same shape is used for summation. σ(·) represents the Sigmoid activation function, which maps the weighted sum to the interval [0,1], so that the output can be used directly as a gating coefficient. The gating weights can be scalars or matrices of the same size as the attention map; when they are close to 0, they rely more on data-driven attention (Attn), and when they are close to 1, they rely more on prior attention. . This represents the cross-attention distribution calculated using the conditional query vector and the visual feature set, corresponding to the result of "performing multi-head cross-attention calculation". It represents the attention distribution constructed from eye-tracking heatmaps, ROI regions, or other interpretation priors, used to characterize the explicit areas of attention of the operator on the image. The fused attention distribution is represented by the gating coefficient. The linear interpolation data-driven attention and prior attention are used as weights for the subsequent generation of the interpretation intent vector.

[0062] This represents the total loss of the interpretation and reconstruction module, used to constrain various output tasks related to image interpretation. This represents the loss for predicting view categories, such as the classification cross-entropy of labels like parasternal long-axis view and apical four-chamber view. This represents the loss for structural visibility assessment, used to supervise the model in determining whether the target anatomical structure is sufficiently visualized. This represents the loss in predicting lesion suspicion, measuring the difference between the lesion risk score output by the model and the manually labeled score. This represents the loss predicted for the next scanning target direction, such as predicting which direction to move into the intercostal space or which imaging parameter needs to be changed. This represents the multimodal alignment loss, used to constrain the distribution consistency between visual features and multiple source signals such as gaze heatmaps and text cues. KL divergence or similarity loss can be used. This represents the loss due to sequential constraints or finite state machine constraints. It is used to ensure that state transitions are consistent with the actual workflow stages and to avoid state jumps that do not conform to the process. , This represents a non-negative weighting coefficient used to adjust the relative weights of alignment loss and state constraint loss in the total loss.

[0063] To obtain the network parameters for image interpretation, intent reasoning, and scan control during the inference phase, a joint optimization approach can be employed during offline training to simultaneously update the interpretation and reconstruction module, action decision module, and phase state modeling module. A multi-task joint loss function is introduced during training, linearly weighting and summing multiple supervision signals. The overall form can be expressed as: in, To assess reconstruction loss, the consistency between the assessment outputs, such as view category, structural visibility, lesion suspicion level, and target orientation, and the manual annotations is constrained. The behavioral imitation loss is used to measure the deviation between the high-level policy instructions and low-level control instructions generated by the model and the doctor's actual operating trajectory. For the purpose Figure 1 To avoid loss of consistency, internal intent representations should be aligned with clinical reading purposes through explicit intent labels or implicit semantic constraints. This is a multimodal alignment loss used to improve the alignment between attention distribution and eye-tracking gaze heatmap, region of interest mask, and workflow stage annotations; This is the stage state constraint loss, used to ensure that the time series of high-level policy categories conforms to the predefined workflow state machine transition relationships, and to suppress unreasonable state jumps. Coefficient , , , , By adjusting the weights of each loss term on the validation set, a balance can be achieved between interpretation accuracy, action stability, and execution logic consistency. By performing gradient backpropagation on the joint loss function and iteratively updating the network parameters, the quality of intent modeling and action generation can be improved simultaneously during the training phase, enabling stable ultrasonic scanning control capabilities during inference execution in the deployment phase.

[0064] This embodiment can directly introduce comprehensive constraints from the operator's gaze behavior, operational semantics, and workflow stage into the image feature encoding stage, so that the selection of visual features no longer depends solely on the image pixels themselves, but is consistent with the clinical interpretation logic. This reduces ambiguity in action decision-making and improves the fit of subsequent trajectory control with the true interpretation intent in cases of complex anatomical structures, uneven lesion distribution, or fluctuating imaging quality.

[0065] In one embodiment, step S40 above includes: S401, Spatial position coordinate data, attitude direction data and contact force numerical data are separated from the motion state data of the end effector; S402, the sliding window algorithm is used to synchronously extract and standardize the spatial position coordinate data, the attitude direction data and the contact force numerical data according to the length of the preset time window, and the processed data are spliced ​​together in the feature dimension to obtain a multi-dimensional time series matrix as a sequence segment. S403, the sequence fragment is mapped to a high-dimensional feature space using a feature projection layer, and temporal position encoding information is superimposed to generate an embedded feature sequence with time stamps; S404, The embedded feature sequence is input into a causal convolutional neural network, and the embedded feature sequence is convolved using multi-layer dilated causal convolution kernels to generate an encoded hidden state sequence containing local dynamic change features and long-range motion trend features. S405, extract the feature vector corresponding to the current time step from the encoded hidden state sequence, and determine the feature vector as a historical trajectory representation containing motion inertial information.

[0066] In this embodiment, the motion state data originates from the real-time feedback of the end effector during the scanning process. It generally includes the end effector's position coordinates in three-dimensional space, its orientation, and the mechanical feedback acting on the probe. Spatial position coordinate data can be represented using Cartesian coordinates, for example, by using a joint encoder in conjunction with kinematic calculations to obtain the x, y, and z numerical sequences of the end effector in the examination table coordinate system. Orientation data can be represented using Euler angles, rotation vectors, or quaternions to represent the orientation change of the end effector probe, expanding the orientation parameters at each time point into a fixed-dimensional vector. Contact force data can be collected by force sensors mounted on the end effector or probe gripping mechanism, including force components along the normal and tangential directions, or six-dimensional force / torque data, which, after sampling, forms a scalar or vector sequence that varies over time. By aligning the position, orientation, and force channels on the time axis, a multi-channel time-series signal can be obtained, completely recording the motion trajectory and mechanical action of the scanning actuator over a period of time.

[0067] To extract local motion processes from continuous time series, windowing needs to be performed on the time axis. The preset time window can be set to a certain number of sampling points based on the clinical operation rhythm and control frequency, for example, covering the most recent second- or sub-second-level motion trajectory. The sliding window algorithm, in its implementation, sets the window length and sliding step size, simultaneously providing start and end indices for spatial coordinate data, attitude orientation data, and contact force numerical data. This ensures that each segment includes all three types of data within the same time span, achieving synchronous segmentation. To eliminate differences in the dimensions and numerical ranges of different physical quantities, standardization processing needs to be performed on each segment of data. This can be achieved by subtracting the mean and then dividing by the standard deviation, or by linear scaling based on empirical upper and lower limits to constrain the values ​​within a uniform range. After standardization, the spatial coordinates, attitude orientation, and contact force values ​​at the same time point are concatenated along the feature dimension to obtain a multidimensional vector. Arranging this vector sequentially along the time axis forms a multidimensional time series matrix. This matrix contains complete state information for each time point within the window while preserving the temporal order, and is used as a sequence segment.

[0068] When sequence fragments are directly used for subsequent modeling, their dimensionality may not match the downstream network structure, or their expressive power may be insufficient. Therefore, a feature projection layer is used to map the sequence fragments to a high-dimensional feature space. The feature projection layer can employ a one-dimensional convolutional layer or a time-step fully connected mapping, transforming the original state vector at each time step into a higher-dimensional feature vector. During this process, state components from different sources can be weighted and combined, allowing the mapped vector to simultaneously encode the coupling relationships of position, pose, and force. To enable the network to distinguish different time positions when processing time series, temporal position encoding information needs to be superimposed on the mapping result. Temporal position encoding information can be generated in a fixed sine / cosine function form or implemented using a learnable vector table, assigning a position vector to each time index and then adding or concatenating it with the mapping vector of the corresponding time step. The feature sequence after superimposing position encoding is considered an embedded feature sequence, maintaining its original order in the time dimension and achieving a unified mapping in the feature dimension.

[0069] Embedded feature sequences are fed into a causal convolutional neural network for temporal modeling. The causal convolutional structure ensures that future information is not introduced into the inference scenario by limiting the receptive field of the convolutional kernel to only cover the current time point and its historical time points, without referencing features from future time points. Convolution operations can use a one-dimensional convolutional kernel sliding along the time axis, with each operation using the embedding vectors from several historical time steps to calculate the output of the current time step. To cover a longer time span within a finite number of layers, a dilation factor is introduced, expanding the sampling interval of the convolutional kernel on the time axis to a fixed stride. This allows the same convolutional kernel to read input across multiple time points in the time dimension, forming dilated causal convolution. By concatenating multiple layers of dilated causal convolutional kernels in the network and appropriately setting the dilation factor of each layer, the network can focus on detailed changes within a short time range in convolutional layers closer to the input layer, and gradually expand the receptive field in deeper convolutional layers, thereby simultaneously capturing local dynamic changes and long-term motion trends. The outputs of the convolutional layer are arranged in chronological order on the time axis, which yields the encoded hidden state sequence. Each time position corresponds to a high-dimensional hidden state vector, which has taken into account the trajectory shape, attitude adjustment and contact force changes over a period of time within the window.

[0070] Historical trajectory representation focuses on the inertial trend at the current control moment. Therefore, the hidden state vector corresponding to the current time step is selected from the encoded hidden state sequence as the representative. The current time step can be defined as the time index at the end of the window, or a relative position within the window can be selected according to the specific application, as long as the position is aligned with the current control moment in time. The extracted hidden state vector is structurally consistent with the network output and numerically contains both local dynamic change information near the current moment and long-range motion trend information within the window, such as probe movement direction, speed changes, and the trend of continuous pressure application or decompression. This hidden state vector is defined as the historical trajectory representation, that is, a fixed-dimensional vector is used to represent the motion inertial information of a short period of time before the current moment in the feature space, providing motion history-related input for subsequent intention-based control decisions.

[0071] Through the above steps, this embodiment can centrally express the motion trajectory, attitude adjustment, and contact force change trend of the scanning actuator in a fixed-length vector over a recent period of time. This provides the subsequent decision-making module with input containing motion inertia information, enabling the control logic to take into account both the current state and short-term motion history when generating actions, thereby reducing the risk of trajectory jitter and sudden operation changes.

[0072] In one embodiment, step S50 above includes: S501, using the determined intent vector as a decoding condition, perform a cross-modal feature fusion operation on the historical trajectory representation to generate action decision features, including: S502, the historical trajectory representation is mapped to a query vector using a query projection layer, and the interpretation intent vector is mapped to a key vector and a value vector using a key-value projection layer respectively. S503, the matrix multiplication operation unit is used to calculate the dot product similarity between the query vector and the key vector, and the dot product similarity is divided by a preset scaling factor to generate an original relevance score that represents the degree of association between the motion state and the judgment logic. S504, The original correlation scores are normalized using the Softmax activation function to generate cross-modal attention weights; S505, perform weighted aggregation calculation on the value vector according to the cross-modal attention weight to generate intention guidance features that are adapted to the historical trajectory representation; S506, perform a residual connection operation on the intention guidance feature and the historical trajectory representation, and use the layer normalization module to normalize the result after the residual connection operation to obtain the action decision feature.

[0073] In this embodiment, using the interpretation intent vector and historical trajectory representation as input, a series of constrained linear transformations and attention operations are used to transform the human interpretation intent into a control decision expression that aligns with the current motion inertia. The interpretation intent vector can be understood as a high-dimensional representation provided by the upstream module after aggregating image information and task interaction information, which includes encoding of interpretation logic such as the current anatomical region, suspicious lesions, and acoustic window optimization requirements. The historical trajectory representation comes from the motion state encoding within the previous time window, including motion inertia information such as the probe's movement trend in space, posture adjustment rhythm, and changes in contact force. Since the two originate from different space sources, it is difficult to characterize the relationship of "which interpretation intents should be responded to under the current motion trend" if they are directly spliced ​​or simply superimposed. Therefore, a matching relationship is established in a unified vector space and action decision features are generated through cross-modal feature fusion operations.

[0074] The query projection layer performs the operation of mapping historical trajectory representations to query vectors. Specifically, a set of learnable weight matrices and bias parameters can be assigned to the historical trajectory representation h, treating h as a one-dimensional row vector. A linear transformation yields a fixed-dimensional query vector q. This transformation can be deployed as a single-layer fully connected mapping or directly converted to a matrix multiplication and addition unit in the hardware implementation, ensuring consistent operational semantics across different hardware platforms. Through this mapping, historical trajectory information is encoded into the attention space, making the query vector comparable to the key vector in subsequent similarity calculations. The key-value projection layer processes the interpretation intent vector z, generating a key vector k and a value vector v through two different sets of linear mapping parameters. This defines the "index representation for matching" and the "content representation for output" respectively in the same interpretation space. The key vector retains the components in the interpretation intent used to distinguish different intent emphases, while the value vector retains information that needs to be injected into the decision representation during fusion, such as bias towards sliding along the intercostal path or preferential coverage of a certain organ boundary.

[0075] The matrix multiplication unit performs a dot product operation on the query vector and key vectors to calculate their similarity in the attention space. In implementation, the query vector can be represented as a single-row multidimensional vector, and multiple key vectors can be concatenated column-wise into a matrix. The similarity components for multiple key vectors are obtained simultaneously through vector-matrix multiplication. To control the numerical scale and avoid excessively large dot product results in high-dimensional space leading to unstable gradient updates, the similarity result is divided by a pre-defined scaling factor. This scaling factor can be the square root of the key vector dimension or an empirical value based on offline statistical results. After scaling, the original relevance score is obtained. This score numerically reflects the degree of matching between the current motion state and different interpretation intent components, and semantically represents "which type of interpretation instruction is more likely to be responded to under the current trajectory inertia."

[0076] The Softmax activation function normalizes the original relevance scores, converting a set of arbitrary real numbers into a non-negative weight vector that sums to one. Numerically, this normalization is achieved through exponential operations and a normalization factor; semantically, it corresponds to a probabilistic characterization of the importance of different interpretive intent components. For relatively high original relevance scores, the Softmax output approaches one while other components approach zero, indicating that the motion state is more biased towards a particular interpretive intent. For scores that are similar, the output weights are more balanced, indicating that multiple interpretive intents need to be considered simultaneously under the current motion trend. This cross-modal attention weight is directly used as a coefficient in subsequent weighted aggregation to adjust the contribution ratio of different value vectors to the final fusion result.

[0077] Weighted aggregation computation uses cross-modal attention weights as coefficients to linearly combine a set of value vectors to generate intent-guided features. In implementation, multiple value vectors can be organized into a matrix column-wise, and attention weight vectors can be used as a set of scalar coefficients. The aggregation result is obtained through matrix multiplication or column-wise multiplication and addition operations. During aggregation, the intent content carried by each value vector participates in the synthesis according to its corresponding weight. The aggregated vector is numerically a convex combination of multiple value vectors, and semantically represents an "intent expression filtered and reconstructed for the current motion state." This intent-guided feature is no longer a simple copy of the original intent vector, but a weighted expression that has been combined with query information from historical trajectory representations, making it more suitable for direct input into the control decision module.

[0078] The residual connection operation adds the intent-guided features and historical trajectory representations element-wise along the same feature dimension. Before addition, a linear mapping is used to align one of the elements to ensure consistent vector length. The introduction of the residual structure allows the fusion result to retain motion inertia information while adding corrections from the interpreted intent, avoiding the loss of original trajectory information due to complete reliance on attention output. This ensures that the action decision expression both follows the continuity of physical motion and reflects the adjustment needs at the cognitive level. The layer normalization module then normalizes the mean and variance of the residual connection result, subtracting the mean of the current sample at that layer from each feature dimension and dividing by the standard deviation. Learnable scaling and offset parameters can also be added to restore an appropriate numerical scale. The normalization operation remains consistent during training and inference, which can reduce distribution differences between different samples and internal covariance shifts, improving the sensitivity and stability of subsequent networks to action decision features. The vector obtained after residual connection and layer normalization is the action decision feature, which combines trajectory inertia representation and intent-guided information in its numerical structure. It can be directly used as input for generating high-level policy instructions and low-level control instructions in downstream modules.

[0079] This embodiment can simultaneously encode the matching relationship between the current motion inertia and the human interpretation intention in a compact representation. Compared with simple splicing or linear superposition, it is easier to highlight the intention component that truly needs to be responded to in the current trajectory state. This ensures that subsequent control decisions will not deviate from the historical motion trend when adjusting the trajectory, and can also selectively deflect the doctor's focus area and operational intention, thereby improving the continuity, relevance and interpretability of the scanning action.

[0080] In one embodiment, step S60 above includes: S601, the action decision features are transmitted in parallel to the strategy classification branch and motion regression branch set inside the hierarchical decoding module; S602, the fully connected classification layer in the strategy classification branch is used to map and calculate the action decision features to obtain the operation mode probability distribution, and the discrete strategy category is determined from the operation mode probability distribution according to the maximum probability principle to obtain the high-level strategy instruction; S603, using the multilayer perceptron network in the motion regression branch to perform nonlinear regression calculation on the action decision features, and outputting a prediction vector containing spatial pose change and contact force control. S604, combine the spatial pose change amount and the contact force control amount into continuous motion parameters, and encapsulate the continuous motion parameters into data to generate low-level control commands.

[0081] In this embodiment, using action decision features as input, the same vector is simultaneously mapped into discrete signals for strategy decision-making and continuous signals for motion control through the parallel structure within the hierarchical decoding module. The action decision features originate from the preceding fusion stage and have already numerically integrated the intent and historical motion inertia, containing comprehensive information on what scanning action should be performed at the current moment and the appropriate force and pose adjustment method. To avoid target conflicts caused by mixing classification decisions and regression control in a single channel, the hierarchical decoding module sets up a strategy classification branch and a motion regression branch. Through a parallel transmission mechanism, the action decision features are copied into two data streams and input into the two branches respectively, ensuring that the classification and regression tasks are independent in parameter space and synchronized in time. This guarantees that the high-level strategy instructions and low-level control instructions generated at the same time originate from the same source.

[0082] The policy classification branch transforms action decision features into an operation mode probability distribution through a fully connected classification layer. Specifically, the action decision features first undergo a linear transformation, mapping the original feature dimensions to an output dimension matching the number of predefined policy categories. This linear transformation consists of a weight matrix and a bias vector, and the influence weights of different feature components on each policy category are learned through training. After the linear output, a normalization operation can be configured. For example, an exponential operation can be performed on the output of each policy category, followed by sum-based normalization, resulting in an operation mode probability distribution with values ​​ranging from zero to one and a sum of one. Each component in this probability distribution corresponds to a candidate policy category, such as different scanning modes, different action stages, or different control modes. The maximum probability principle selects the index corresponding to the component with the largest value in this distribution, transforming the continuous probability expression into a discrete policy category, thus avoiding control logic conflicts caused by simultaneous activation of multiple policies. In this processing stage, the discrete policy categories are further encapsulated into high-level policy instructions. These instructions can include information such as the category index, the marker of the current decision moment, and the policy duration estimate, providing clear policy labels for subsequent control links.

[0083] The motion regression branch employs a multilayer perceptron network (MLB) to perform nonlinear regression calculations on action decision features, outputting a prediction vector containing spatial pose changes and contact force control quantities. The MLB consists of several linear layers and nonlinear activation functions stacked sequentially. The linear layers adjust the feature dimensions and feature combination methods through matrix multiplication and addition operations, while the nonlinear activation functions, such as rectification or smoothing functions, introduce nonlinear transformation capabilities, enabling the network to fit complex mapping relationships. The input action decision features undergo multiple linear and nonlinear transformations within this network, gradually projecting from a high-dimensional abstract space to a vector space with clear physical meaning. The final output prediction vector contains two parts: spatial pose changes and contact force control quantities. Spatial pose changes can be expressed as three-dimensional or six-dimensional vectors, such as translational and rotational increments, while contact force control quantities can be expressed as one-dimensional or multi-dimensional scalars, such as target pressure in the probe normal direction or the resultant force of a target in multiple directions. Through the training process, different components are automatically aligned to their corresponding physical quantities, allowing the prediction vector to directly serve as the source of pose and force targets.

[0084] Spatial pose change and contact force control variables numerically constrain the scanning robot's end effector's motion trend in space and its contact relationship with the object being inspected, respectively. Therefore, they need to be combined in a unified structure for unified management and distribution. Continuous motion parameters are the result of this combination. In implementation, the spatial pose change vector and the contact force control vector can be concatenated along their dimensional directions to form a long vector containing position increments, attitude increments, and force control targets. Alternatively, a structured data type can be defined to store the two types of values ​​as key-value pairs to adapt to different software architectures. After generation, continuous motion parameters require data encapsulation to meet the requirements of communication protocols, cache structures, or hardware interfaces. The original floating-point or fixed-point numbers are arranged according to a predefined field order, and necessary timestamps, channel identifiers, or verification information are added to ultimately form low-level control instructions. These low-level control instructions are compact in expression for rapid transmission on the control bus while retaining sufficient pose and force control resolution to support fine-grained trajectory adjustments. Through this hierarchical decoding structure, motion decision features are simultaneously interpreted as high-level policy instructions and low-level control instructions, providing synchronized and consistent instruction pairs for subsequent execution modules.

[0085] This embodiment transmits action decision features in parallel to the strategy classification branch and motion regression branch within the hierarchical decoding module. At the same time, it generates discrete strategy categories derived from the probability distribution of operation modes and spatial pose change and contact force control quantities obtained from multilayer perceptron regression. These are further combined to form continuous motion parameters and encapsulated as low-level control commands. Without introducing target conflicts, the abstract decision expression can be split into two complementary types of control signals. This allows the scanning robot to clearly understand the current operation mode at a high level and obtain continuous and physically consistent pose and force control targets at a low level. This improves the decision clarity and execution precision of scanning actions and reduces coupling interference between strategy switching and trajectory adjustment.

[0086] In one embodiment, a scanning robot control device driven by intent determination is provided, which corresponds one-to-one with the scanning robot control method driven by intent determination in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the intent-driven scanning robot control device of the present invention. The module includes a data acquisition module 10, a visual feature extraction module 20, an intent reasoning module 30, a trajectory encoding module 40, a feature fusion module 50, a hierarchical decoding module 60, and an execution control module 70. Detailed descriptions of each functional module are as follows: The data acquisition module 10 is used to acquire in real time the scan image sequence, motion status data of the end effector, and task interaction reference data generated during the scanning process; The visual feature extraction module 20 is used to perform convolution processing on the scanned image sequence using a visual feature extraction network to generate a set of visual features containing multi-resolution information. The intent reasoning module 30 is used to map the task interaction reference data into a condition query vector, and use the condition query vector as an attention guidance signal to perform attention reasoning calculation on the visual feature set to generate a judgment intent vector representing the judgment decision logic. The trajectory encoding module 40 is used to extract the sequence segments of the motion state data within a preset time window, and to encode the sequence segments with temporal features to generate a historical trajectory representation containing motion inertial information. Feature fusion module 50 is used to perform cross-modal feature fusion operation on the historical trajectory representation using the interpretation intent vector as a decoding condition to generate action decision features; The hierarchical decoding module 60 is used to input the action decision features into the hierarchical decoding module for parallel prediction, and generate high-level policy instructions containing discrete policy categories and low-level control instructions containing continuous motion parameters. The execution control module 70 is used to synthesize a drive signal based on the high-level strategy instructions and the low-level control instructions to control the scanning robot to perform corresponding scanning operations.

[0087] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of a server-side control method for a scanning robot driven by intent interpretation.

[0088] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements client-side functions or steps of a scan robot control method driven by intent interpretation.

Claims

1. A scanning robot control method driven by intent interpretation, characterized in that, Includes the following steps: Real-time acquisition of scan image sequences, end effector motion status data, and task interaction reference data generated during the scanning process; The scanned image sequence is convolved using a visual feature extraction network to generate a set of visual features containing multi-resolution information. The task interaction reference data is mapped to a conditional query vector, and the conditional query vector is used as an attention guidance signal to perform attention inference calculation on the visual feature set to generate a judgment intent vector that represents the judgment decision logic. Extract sequence segments of the motion state data within a preset time window, and encode the sequence segments with temporal features to generate a historical trajectory representation containing motion inertial information; Using the interpretation intent vector as a decoding condition, a cross-modal feature fusion operation is performed on the historical trajectory representation to generate action decision features; The action decision features are input into the hierarchical decoding module for parallel prediction to generate high-level policy instructions containing discrete policy categories and low-level control instructions containing continuous motion parameters. The high-level strategy instructions and the low-level control instructions are combined to form a drive signal, which controls the scanning robot to perform the corresponding scanning operation.

2. The intention-driven scanning robot control method as described in claim 1, characterized in that, Real-time acquisition of scan image sequences, end effector motion state data, and task interaction reference data generated during the scanning process, including: A communication connection with the scanning imaging device is established using a video acquisition interface, and video stream frame images output by the scanning imaging device are acquired in real time through the video acquisition interface at a preset sampling frequency. The video stream frame images are arranged in chronological order of acquisition time to generate a scanned image sequence. The robot control system is connected via a data communication interface, and the six-degree-of-freedom pose parameters fed back by the robot control system are read in real time through the data communication interface. The real-time contact force parameters are collected by a six-dimensional force sensor installed on the end effector. The six-degree-of-freedom pose parameters and the real-time contact force parameters are combined to generate motion state data of the end effector. The screen gaze coordinates when the operator gazes at the display terminal are captured by an eye-tracking sensor, and the operation command signal and workflow stage identifier input by the operator are received by the human-computer interaction interface. The screen gaze coordinates, the operation command signal and the workflow stage identifier are encapsulated and processed to generate task interaction reference data. The image frame timestamps of the scanned image sequence, the state timestamps of the motion state data of the end effector, and the interaction timestamps of the task interaction reference data are obtained respectively, and the image frame timestamps are set as the reference time axis. A linear interpolation algorithm is used to resample the motion state data of the end effector and the task interaction reference data based on the state timestamp and the interaction timestamp, and the motion state data of the end effector and the task interaction reference data are mapped onto the reference time axis.

3. The intention-driven scanning robot control method as described in claim 1, characterized in that, The scanned image sequence is convolved using a visual feature extraction network to generate a set of visual features containing multi-resolution information, including: The image frames in the scanned image sequence are input into the backbone convolution module of the visual feature extraction network. The backbone convolution module performs a stepwise downsampling convolution operation on the image frames to generate multiple low-level feature maps with different spatial dimensions and channel dimensions. Using the feature pyramid module of the visual feature extraction network, the high-level semantic feature map in the bottom feature map is selected for upsampling processing, and the spatial resolution of the high-level semantic feature map is enlarged to match the low-level geometric feature map in the bottom feature map. The channel dimension of the low-level geometric feature map is adjusted by connecting the horizontal channels, and the adjusted low-level geometric feature map is superimposed with the upsampled high-level semantic feature map to generate a fused feature map that integrates semantic information and spatial details. The fused feature map is smoothed by using convolution kernels to eliminate upsampling aliasing, and multiple pyramid feature maps of different scales are output. The multiple pyramid feature maps of different scales are then combined to form a visual feature set containing multi-resolution information.

4. The intention-driven scanning robot control method as described in claim 1, characterized in that, The task interaction reference data is mapped to a conditional query vector, and the conditional query vector is used as an attention guidance signal to perform attention inference calculation on the visual feature set, generating a judgment intent vector representing the judgment decision logic, including: Analyze the task interaction reference data to extract screen gaze point coordinates, operation command signals, and workflow stage identifiers; The screen gaze point coordinates are mapped to spatial position embedding vectors using a position encoder, and the operation command signal and the workflow stage identifier are mapped to semantic command embedding vectors and stage state embedding vectors respectively using a semantic embedding layer. The spatial location embedding vector, the semantic instruction embedding vector, and the stage state embedding vector are subjected to feature concatenation processing to obtain the conditional query vector; Using the conditional query vector as the query term and mapping the visual feature set as a key value term, multi-head cross-attention calculation is performed. The visual feature set is weighted and aggregated by calculating the matching degree between the query term and the key value term to obtain intermediate interpretation features. A gating network is used to calculate prior adjustment weights based on the stage state embedding vector, and the intermediate interpretation features are weighted and corrected using the prior adjustment weights to generate a interpretation intent vector.

5. The intention-driven scanning robot control method as described in claim 1, characterized in that, Extracting sequence segments of the motion state data within a preset time window, and encoding the sequence segments using temporal features to generate a historical trajectory representation containing motion inertial information, including: Spatial position coordinate data, attitude direction data, and contact force numerical data are extracted from the motion state data of the end effector; The sliding window algorithm is used to synchronously extract and standardize the spatial position coordinate data, the attitude direction data, and the contact force numerical data according to the length of the preset time window. The processed data are then spliced ​​together along the feature dimension to obtain a multidimensional time series matrix as a sequence segment. The sequence fragments are mapped to a high-dimensional feature space using a feature projection layer, and temporal position encoding information is superimposed to generate an embedded feature sequence with time stamps; The embedded feature sequence is input into a causal convolutional neural network, and multi-layer dilated causal convolution kernels are used to perform convolution operations on the embedded feature sequence to generate an encoded hidden state sequence containing local dynamic change features and long-range motion trend features. The feature vector corresponding to the current time step is extracted from the encoded hidden state sequence, and the feature vector is determined as a historical trajectory representation containing motion inertial information.

6. The intention-driven scanning robot control method as described in claim 1, characterized in that, Using the determined intent vector as a decoding condition, a cross-modal feature fusion operation is performed on the historical trajectory representation to generate action decision features, including: The historical trajectory representation is mapped to a query vector using a query projection layer, and the interpretation intent vector is mapped to a key vector and a value vector using a key-value projection layer, respectively. The dot product similarity between the query vector and the key vector is calculated using a matrix multiplication operation unit, and the dot product similarity is divided by a preset scaling factor to generate an original relevance score that represents the degree of association between the motion state and the judgment logic. The original relevance scores are normalized using the Softmax activation function to generate cross-modal attention weights; The value vector is weighted and aggregated according to the cross-modal attention weights to generate intent guidance features that are adapted to the historical trajectory representation; A residual connection operation is performed on the intention guidance features and the historical trajectory representation, and the result after the residual connection operation is normalized using a layer normalization module to obtain the action decision features.

7. The intention-driven scanning robot control method as described in claim 1, characterized in that, The action decision features are input into the hierarchical decoding module for parallel prediction, generating high-level policy instructions containing discrete policy categories and low-level control instructions containing continuous motion parameters, including: The action decision features are transmitted in parallel to the strategy classification branch and motion regression branch set inside the hierarchical decoding module; The action decision features are mapped and calculated using the fully connected classification layer in the strategy classification branch to obtain the operation mode probability distribution. The discrete strategy category is then determined from the operation mode probability distribution according to the maximum probability principle to obtain the high-level strategy instruction. The motion regression branch uses a multilayer perceptron network to perform nonlinear regression on the action decision features, and outputs a prediction vector containing spatial pose change and contact force control. The spatial pose change and the contact force control quantity are combined into continuous motion parameters, and the continuous motion parameters are encapsulated to generate low-level control commands.

8. A scanning robot control device driven by interpretation of intent, characterized in that, The intent-driven scanning robot control device includes: The data acquisition module is used to acquire in real time the scan image sequence, motion status data of the end effector, and task interaction reference data generated during the scanning process; The visual feature extraction module is used to perform convolution processing on the scanned image sequence using a visual feature extraction network to generate a set of visual features containing multi-resolution information. The intent reasoning module is used to map the task interaction reference data into a condition query vector, and use the condition query vector as an attention guidance signal to perform attention reasoning calculation on the visual feature set to generate a judgment intent vector that represents the judgment decision logic. The trajectory encoding module is used to extract the sequence segments of the motion state data within a preset time window, and to encode the sequence segments with temporal features to generate a historical trajectory representation containing motion inertial information. The feature fusion module is used to perform cross-modal feature fusion operation on the historical trajectory representation using the interpretation intent vector as a decoding condition to generate action decision features; The hierarchical decoding module is used to input the action decision features into the hierarchical decoding module for parallel prediction, and generate high-level policy instructions containing discrete policy categories and low-level control instructions containing continuous motion parameters. The execution control module is used to synthesize a drive signal based on the high-level strategy instructions and the low-level control instructions, and control the scanning robot to perform corresponding scanning operations.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a readability intent-driven scanning robot control program stored in the memory and executable on the processor. When executed by the processor, the readability intent-driven scanning robot control program implements the steps of the readability intent-driven scanning robot control method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a scanning robot control program driven by the intent to be interpreted, which, when executed by a processor, implements the steps of the scanning robot control method driven by the intent to be interpreted as described in any one of claims 1-7.

Images