Mechanical arm track segmentation method based on Z-axis state transition and speed trough

Abstract

The invention discloses a mechanical arm track segmentation method based on Z-axis state transition and speed troughs. According to the method, a Z-axis height sequence and a speed vector sequence are extracted from a motion trail position sequence of an end effector of the mechanical arm. Firstly, features of a Z-axis height sequence are counted, a relative threshold value is set, Z-axis height is subjected to state division, and state conversion points are searched; and the moment of conversion from the high-order state or the intermediate state to the low-order state is paired with the nearest moment of conversion from the low-order state to the high-order state or the intermediate state to form a candidate interval. Then selecting an interval with the longest time span from all candidate intervals as a rough candidate interval; and then speed trough detection is conducted on the speed size sequence in the rough candidate interval, speed minimum value points closest to the starting point and the ending point of the interval are found to serve as boundary candidate points of starting and ending, a boundary point pair based on speed troughs is constructed, a mechanical arm track is segmented, and an action track of task execution is obtained.

Description

Technical Field

[0001] This invention belongs to the field of automation control technology and relates to robotic arm trajectory analysis, specifically a method for automatic segmentation of robotic arm motion trajectory based on Z-axis state transition and velocity trough. Background Technology

[0002] Robotic arm teaching is a crucial method for achieving rapid programming and skill transfer in robotic arms. During robotic arm teaching, operators guide the robotic arm to complete a target task through teleoperation, drag-and-drop teaching, or visual tracking. The system automatically records the complete motion trajectory, including Cartesian spatial position and posture. However, the complete trajectory often contains a large amount of redundant information, such as preparation phases before the action, adjustment phases after the action, and irrelevant segments like waiting pauses. Accurately extracting the truly effective task execution segments from the complete teaching trajectory has become a key technical challenge in the field of robotic arm teaching.

[0003] Prior art 1 (CN117921669A) discloses a robotic arm trajectory segmentation method based on curvature analysis. This method first calculates the velocity and acceleration vectors for each position point in the trajectory point sequence to represent the instantaneous motion direction and rate of velocity change. Then, it calculates the local curvature value using a curvature formula to represent the degree of trajectory bending. Finally, it calculates the cosine of the angle between adjacent direction vectors; when the cosine of the angle is less than a preset threshold, it is determined to be a significant change in direction and marked as a trajectory segmentation point. This method utilizes the geometric features of curvature and direction changes for trajectory segmentation and has a certain degree of universality. However, curvature calculation involves the cross product of velocity and acceleration, which is equivalent to performing a second-order numerical differentiation on the position signal. This process is quite sensitive to trajectory noise and is prone to producing false curvature peaks on actual trajectories containing sensor noise. Furthermore, a fixed directional change threshold lacks adaptability to different motion speeds and scales, making it difficult to simultaneously adapt to fast motion and slow, precise operations. In addition, curvature, as a purely geometric feature, does not directly correspond to the semantic boundary of the task. Especially in grasping tasks, the trajectory curvature during the approach and grasping phases may change continuously without obvious abrupt changes, leading to a deviation between the segmentation results and the task semantics.

[0004] Existing technology 2 (CN120080315A) proposes a trajectory segmentation method based on Dynamic Motion Element (DMP). This method is based on the assumption that robotic arm operations typically occur near the lowest point of the Z-axis. First, it detects the global lowest point of the Z-axis height in the complete trajectory sequence. Then, it expands a fixed window forward and backward around this point, extracting trajectory segments within the expanded time window and fitting a DMP model. While this method has some physical intuition, in complex tasks, the lowest point may appear at any stage of the action rather than at a critical semantic boundary. Furthermore, the duration of actions varies significantly across different task types, causing the expanded time window to truncate the complete action or contain excessive redundant information, failing to accurately cover effective action segments. This method is particularly ineffective for complex tasks involving multiple contacts or where Z-axis changes are not significant.

[0005] In summary, the core problems faced by current trajectory segmentation methods include: insufficient adaptability, as fixed parameters and thresholds are difficult to adapt to different task scenarios and environmental changes; limited robustness, being sensitive to sensor noise and numerical differential errors; limited accuracy, with a lack of fine-grained correction steps leading to boundary positioning deviations; and a lack of semantic understanding, as methods based on low-level geometric features struggle to capture the high-level semantic structure of the task. Therefore, a robotic arm trajectory segmentation method that combines high accuracy, strong robustness, and adaptability is needed, capable of reliably and stably extracting effective task action segments in complex real-world environments, providing accurate reference information for subsequent trajectory learning, skill generalization, and action optimization. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a robotic arm trajectory segmentation method based on Z-axis state transition and velocity valley. Combining coarse positioning via Z-axis state transition with fine correction via velocity valley, this two-stage adaptive segmentation method achieves high-precision and robust trajectory segmentation under different task scenarios, robotic arm platforms, and noise levels. It can be widely applied to robotic arm teaching and learning, skill representation and recognition, motion trajectory optimization, human-machine collaborative teaching, industrial automation programming, and other scenarios, solving the problems of low accuracy, poor adaptability, and insufficient robustness in existing trajectory segmentation methods.

[0007] A robotic arm trajectory segmentation method based on Z-axis state transition and velocity trough includes the following steps:

[0008] Step 1: Trajectory Data Acquisition

[0009] Collect the complete motion trajectory of the robotic arm's end effector in Cartesian space and record the timestamp sequence. With the corresponding position sequence ,in Indicates the total number of frames in the trajectory. , Indicates the first Frame timestamp Indicates the first The three-dimensional position coordinates of the end effector of the frame in Cartesian coordinates. , , These represent the X, Y, and Z axis coordinate components, respectively. From the position sequence Extracting the Z-axis height sequence Used for subsequent state transition analysis.

[0010] Step 2: Z-axis feature extraction

[0011] Because the motion amplitude varies significantly across different tasks, using an absolute threshold can lead to oversensitivity in tasks with small amplitudes and insufficient sensitivity in tasks with large amplitudes. This method addresses this issue by adjusting the Z-axis height sequence. Perform statistical feature calculations, including the global minimum value of the Z-axis height. and global maximum value Global range of Z-axis height and the mean of the Z-axis height sequence and standard deviation .

[0012] These statistical features reflect the overall motion characteristics of the trajectory in the Z-axis direction. By extracting these relative features, a data basis can be provided for subsequent adaptive threshold calculation to achieve adaptive adjustment of the threshold and ensure the applicability of the method in different task scenarios.

[0013] Step 3: Adaptive threshold determination

[0014] Based on the Z-axis statistical features extracted in step 2, the relative low-order threshold is calculated. and relative high threshold Used for state partitioning:

[0015]

[0016]

[0017] in, , ∈(0,1), representing the relative height coefficients of the low-level and high-level thresholds, respectively, are used to define the boundaries between the low-level and high-level states. The relative threshold method allows the threshold to be automatically adjusted according to the motion amplitude of different trajectories, avoiding the poor adaptability of fixed thresholds in different task scenarios. Since the robotic arm typically needs to approach the work plane when performing grasping or placing actions, the Z-axis height will significantly decrease to the low-level state; while during movement or standby phases, the end effector usually remains at a higher position to avoid collisions. Setting a relative low-level threshold can effectively capture the motion execution phase approaching the work plane; setting a relative high-level threshold as the high-level boundary can distinguish obvious high-low transitions, and the setting of intermediate states enhances the method's adaptability to gradual descent or ascent motions.

[0018] Step 4: Z-axis state transition detection

[0019] Based on the relative low threshold determined in step 3 and relative high threshold The elements in the Z-axis height sequence Z are divided into states to form a state sequence. :

[0020]

[0021] Because robotic arms follow a typical "approach-execution-departure" motion pattern when performing actions—that is, descending from a high position to a low position to perform the action, and then rising from a low position back to a high position—the state transition sequence usually presents a pattern of "high / medium → low → high / medium." By detecting this state transition pattern, the rough boundaries of action segments can be effectively identified. Traversing the state sequence... It detects transitions from high / intermediate states to low states (decreasing) and transitions from low states to high / intermediate states (increasing), and defines a set of decreasing transition points. and the set of rising transition points :

[0022]

[0023]

[0024] like or This indicates that there is no obvious transition between high and low states in the trajectory.

[0025] Step 5: Conversion point pairing

[0026] Based on the set of descent transition points detected in step 4 and the set of rising transition points Pairing transition points. For the set of descent transition points. For each descending transition point in the set of its subsequent ascending transition points... Search for the first rising transition point in the middle to perform pairing:

[0027]

[0028] Where min represents taking the minimum value. , Represents the set of descent transition points The first in Each descent transition point and its corresponding timestamp , Represents the set of rising transition points The Middle Each rising transition point and its corresponding timestamp. Represents the set of rising transition points Middle and the first A descent transition point The rising transition point of the pairing.

[0029] Two successful pairings form a candidate interval. Calculate the time span of the interval Since real robotic arm movements typically require a certain execution time, excessively short low-level dwell times often correspond to measurement noise or momentary jitter, rather than actual movement. Furthermore, the execution time of a single atomic movement usually does not exceed a few seconds, while excessively long low-level dwell times may involve multiple movements or abnormal pauses. Therefore, setting a minimum time span threshold is necessary. and maximum time span threshold Only retain those that meet the requirements. Candidate intervals are used as valid pairing intervals. Time span constraints can effectively filter out false transition points, improving the reliability of pairing.

[0030] In trajectories containing multiple brief approach movements, the actual working movement typically corresponds to the longest low-position dwell time. Therefore, among all valid pairing intervals, the interval with the longest time span is selected as the coarse candidate interval. It can prioritize the identification of key action segments in complex trajectories.

[0031] If no transition point is detected in step 4, or no valid pairing interval is detected in step 5, the detection fails, and a relatively low threshold is used. The mark satisfies For all frames, find the longest consecutive low-order segment as the action interval and segment the robot arm trajectory.

[0032] Detection failures typically occur in the following scenarios: (1) the Z-axis movement amplitude of the trajectory is extremely small, resulting in unclear distinction between high and low states; (2) the robotic arm moves repeatedly across multiple height levels, generating complex state transition sequences; (3) the action execution time is extremely short, and the low-position dwell time does not meet the time constraint. In these cases, the robotic arm still needs to approach the work plane to perform the action, so the Z-axis height will still drop to a relatively low position. Using a relatively low-position threshold... Although the segmentation method has lower accuracy, it is more robust and can still provide reasonable segmentation results based on height information when the state transition is not obvious, ensuring the usability of the method in various complex scenarios.

[0033] Step 6: Speed ​​Calculation and Smoothing

[0034] The position sequence obtained in step 1 The forward difference method is used to calculate the velocity magnitude sequence. ,in Indicates the first The velocity value corresponding to the frame.

[0035] Preferably, due to inherent noise in the robotic arm's sensor measurements and the potential for high-frequency vibrations or jitters during the robotic arm's movement, the directly calculated velocity sequence often contains significant instantaneous fluctuations. These fluctuations can generate numerous spurious extreme points in subsequent trough detection. To suppress measurement noise, a Gaussian filter is used to filter the velocity vector sequence. Perform smoothing processing and set the filter standard deviation. A smooth velocity sequence is obtained. .

[0036] Step 7: Velocity trough detection

[0037] Velocity troughs correspond to the momentary standstill or extremely low-speed states of the robotic arm's movement, typically appearing at the actual start and end points of the action. Because the robotic arm needs to decelerate from a moving state to a standstill before starting an action to ensure positioning accuracy, and also needs to accelerate away from a standstill after the action is completed, the velocity curve exhibits a distinct trough characteristic at the boundaries of the actual action. Compared to Z-axis state transition points, velocity troughs more directly reflect changes in motion state, thus providing more accurate boundary positioning. Furthermore, since the coarse candidate intervals obtained in step 5 may contain partial deceleration or acceleration phases at the boundaries, searching for velocity troughs within the candidate intervals can further refine the boundaries and eliminate these transitional phases.

[0038] The coarse candidate interval determined in step 5 Internal velocity vector sequence Perform velocity trough detection. Define the front-end search window. ,in, Indicates the length of the search window. In the foreground search window... Find the point with the minimum velocity within the boundary as a candidate point for the initial boundary. :

[0039]

[0040] Define the backend search window In the backend search window Find the point with the minimum velocity within the boundary as a candidate point for the final boundary. :

[0041]

[0042] This method yields boundary point pairs based on velocity troughs. .

[0043] Step 8: Boundary Constraints

[0044] Step 7 yields boundary point pairs based on velocity troughs. Perform constraint verification. Since velocity trough detection is based on local minimum search, in extreme cases such as excessively short candidate intervals or abnormally flat velocity curves, non-physical situations may occur where the endpoint occurs before the starting point. Ensure boundary monotonicity. If the condition is not met, the endpoint will be reset to the endpoint of the coarse candidate interval, i.e., let This adjustment strategy, while ensuring the rationality of the temporal logic, retains the refined starting point as much as possible. The intervals with boundary constraints are used as the final segmentation boundaries. The trajectory of the robotic arm is segmented to obtain the motion trajectory of the task execution.

[0045] The present invention has the following beneficial effects:

[0046] 1. Due to the significant differences in the Z-axis motion amplitude for different operational tasks, this invention utilizes a relative height coefficient. , The method calculates low-order and high-order thresholds, enabling the thresholds to automatically adjust with the amplitude of trajectory movement. This adapts to the motion scale of different tasks without requiring repeated parameter adjustments for different scenarios, significantly improving the versatility and practicality of the method. It also solves the problem that existing fixed threshold schemes are too sensitive in small-amplitude tasks and insufficient in response to large-amplitude tasks.

[0047] 2. A two-stage strategy is adopted: coarse localization via Z-axis state transition and fine localization via velocity troughs. The first stage uses Z-axis state transition to detect candidate motion intervals, providing robust coarse boundaries. The second stage precisely locates the boundaries within the candidate intervals using velocity troughs, avoiding the noise amplification problem caused by the second-order differential required for curvature calculation in existing technologies. This layered strategy organically combines robustness and accuracy, ensuring reliability in noisy environments while achieving precise boundary localization.

[0048] 3. By employing state transition detection in a single traversal and trough search within a local window, compared to the DMP fitting method requiring iterative optimization or the dynamic programming algorithm requiring global search, the computational overhead of this invention is significantly reduced, with an algorithm complexity of linear O(N). Since it does not involve complex numerical optimization and model fitting processes, this method can run efficiently on ordinary industrial control computers. This is particularly important for online teaching scenarios requiring real-time processing and rapid response, enabling the provision of segmentation results immediately after the operator completes the teaching process. It supports an interactive teaching programming workflow, meeting the real-time feedback requirements during robotic arm teaching. Attached Figure Description

[0049] Figure 1 The flowchart shows a robotic arm trajectory segmentation method based on Z-axis state transition and velocity valley detection.

[0050] Figure 2 The Z-axis height sequence extracted in the example and the detection results of the transformation point are shown.

[0051] Figure 3 The conversion point pairing results are shown in the example.

[0052] Figure 4 This is the velocity sequence segmentation result based on the coarse candidate interval in the example;

[0053] Figure 5 The results of trough detection for the velocity sequence in the example;

[0054] Figure 6 This is the velocity sequence segmentation result based on the velocity valley boundary points in the embodiment;

[0055] Figure 7 This is a schematic diagram of the three-dimensional spatial trajectory sequence of the robotic arm to be segmented in the embodiment.

[0056] Figure 8 This example compares the robotic arm trajectory results obtained by segmentation using different methods. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0058] This invention provides a robotic arm trajectory segmentation method based on Z-axis state transition and velocity trough detection. This method analyzes the motion characteristics of the robotic arm's end effector during marking tasks, automatically identifies and extracts the marking action segments, thereby achieving accurate trajectory segmentation. Figure 1 As shown, the process includes Z-axis state transition detection, transition point pairing, and velocity trough refinement. The specific steps are as follows:

[0059] Step 1: Trajectory Data Acquisition

[0060] The robotic arm's end effector is tracked in real time using inverse kinematics, with a sampling frequency set to 50Hz, to obtain a three-dimensional spatial trajectory sequence. And record the corresponding timestamp sequence. ,in, =220 frames, approximately 4.4 seconds long, containing the complete process of the robotic arm moving from the initial position to the marked position, performing the marking action, and returning to the initial position.

[0061] The 3σ criterion is used to identify and remove data points in the collected raw trajectory data that deviate from the mean by more than three times the standard deviation. Gaussian filter parameters are then set. The trajectory data, after outlier removal, is smoothed and filtered to suppress sensor noise and system jitter. The preprocessed trajectory data retains the original motion characteristics while reducing noise interference.

[0062] Extract the Z-axis height sequence from the preprocessed trajectory data. This is used for subsequent state transition analysis.

[0063] Step 2: Z-axis state transition detection and coarse segmentation

[0064] right Figure 2 The Z-axis height sequence shown Perform statistical characteristic calculations to find the global minimum value of the Z-axis height. Maximum value Global scope .

[0065] Set the relative height coefficient between the low-order threshold and the high-order threshold. , Calculate the relative low-order threshold and relative high threshold :

[0066]

[0067]

[0068] Through relatively low threshold and relative high threshold The Z-axis height sequence is divided into three states to form a state sequence. :

[0069] ①When It was in a high-level state at the time. The green area in the diagram indicates that the robotic arm is in standby or moving phase.

[0070] ②When It is currently in a low-level state. The red area in the diagram indicates that the robotic arm approaches the marked surface to perform an action.

[0071] ③When This is an intermediate state. The yellow area in the diagram represents the transition phase.

[0072] Traversing the state sequence In all adjacent frames, when the state transitions from a high bit or middle bit to a low bit, and When, record as the descent transition point. This indicates that the robotic arm has begun to approach the marked surface; when the state transitions from low to middle or high... and When this point is reached, it is recorded as the upward transition point. This indicates that the robotic arm has left the marked surface.

[0073] Figure 2 In the process, two falling transition points were detected in frames 42 and 95, and two rising transition points were detected in frames 59 and 174.

[0074] A forward matching strategy is used to pair the detected transition points, such as Figure 3 As shown, for each descent transition point Find the first rising transition point in its subsequent frames. This forms a candidate interval. And calculate the time span of that interval. Set minimum time span threshold. Frames and maximum time span threshold Frame, only if it satisfies The pairing of candidate intervals is considered a valid pairing interval.

[0075] In this embodiment, the first candidate interval The time span is 17 frames, the second candidate interval The time span is 79 frames, all of which are valid pairing intervals. The valid pairing interval with the longest duration was ultimately selected as the rough candidate interval. .

[0076] Step 3: Velocity trough detection and boundary refinement

[0077] Based on the velocity changes, the boundaries of the rough candidate intervals obtained in step 2 are refined. First, the three-dimensional trajectory sequence is calculated. Displacement vector between adjacent frames :

[0078]

[0079] Then the first Instantaneous velocity of a frame for:

[0080]

[0081]

[0082] in, This is the time interval between adjacent frames. (When the sampling frequency...) When constant, ,but:

[0083]

[0084] Use coarse candidate intervals for velocity sequences Mark, such as Figure 4 As shown, the original velocity sequence remains at a high level outside the coarse candidate interval, while it decreases significantly within the coarse candidate interval. This is consistent with the physical characteristic that a robotic arm needs to reduce its motion speed when performing fine operations.

[0085] The velocity sequence was calculated. Then, a Gaussian filter is used for smoothing, with the filter parameters... A smooth velocity sequence is obtained. For smooth velocity sequences Establish lengths at the front and back ends of the coarse candidate interval respectively. The search window performs trough detection, and the range of the front-end search window is... The backend search window range is The minimum velocity point is searched in both the front-end and back-end search windows to serve as candidate starting boundary points. With the candidate point of the end boundary ,Right now:

[0086]

[0087]

[0088] like Figure 5As shown, in this embodiment, the front-end search window is [95, 135], the detected starting boundary candidate point is located in the 99th frame, and the speed is 0.038m / s. The back-end search window is [134, 174], the detected ending boundary candidate point is located in the 167th frame, and the speed is 0.043m / s.

[0089] Using boundary point pairs based on velocity troughs For smooth velocity sequences Mark, such as Figure 6 As shown, the average velocity is 0.665 m / s before the initial boundary candidate point (frames 0-99), 0.117 m / s between the two boundary candidate points (frames 99-167), and 0.656 m / s after the ending boundary candidate point (frames 167-220). This significant velocity difference confirms the effectiveness of using velocity troughs as motion boundary points. Furthermore, the velocities at the two boundary candidate points are significantly lower than the high-speed movement phases before and after marking, indicating that the robotic arm exhibits a clear stationary or extremely low-speed state at the start and end of the motion. These troughs accurately identify the actual boundaries of the marking motion.

[0090] Compared to the coarse candidate intervals, at the initial boundary, the coarse segmentation point is located at frame 95, with a Z-axis height of 0.383m and a speed of 0.186m / s. At this point, the robotic arm has just entered the low-position state but its speed is still relatively high and has not yet fully stabilized. The initial boundary based on the speed trough is located at frame 99, with a Z-axis height of 0.281m and a speed of 0.038m / s. At this point, the robotic arm has approached the marked surface and is almost stationary, more accurately corresponding to the actual start time of the marking action. At the end boundary, the coarse segmentation point is located at frame 174, with a Z-axis height of 0.406m and a speed of 0.143m / s. At this point, the robotic arm has begun to rise rapidly. The end boundary based on the speed trough is located at frame 167, with a Z-axis height of 0.263m and a speed of 0.043m / s. At this point, the robotic arm has just completed the marking action and has begun to slowly move away, more accurately corresponding to the actual end time of the marking action.

[0091] Using interval pairs formed by the initial boundary candidate points and the end boundary candidate points Figure 7 The three-dimensional spatial trajectory sequence shown Segmentation is performed and compared with segmentation methods based on Z-axis height, such as... Figure 8As shown, (a) is the segmentation result based on Z-axis height, and (b) is the segmentation result of the proposed method. The 3D spatial trajectory sequence before segmentation contains 220 frames of data, of which a large amount of movement processes and standby states are redundant information. The trajectory segmented based on Z-axis height contains 79 frames. Although it has captured the main marked actions, there are deviations in the start and end boundaries. The trajectory segment segmented by the proposed method contains 68 frames of data. The start boundary is adjusted backward by 4 frames, and the end boundary is adjusted forward by 7 frames, reducing the total length by 11 frames. This eliminates the transition and jitter segments near the action boundaries, making the segmentation boundaries closer to the start and end times of the actual actions. Only the core execution process of the marked actions is retained, effectively removing redundant movement segments before and after.

[0092] In summary, this method achieves coarse localization through Z-axis state transition detection and refines boundaries through velocity trough detection, effectively solving the problems of boundary ambiguity and reliance on manual annotation inherent in traditional methods. The segmentation results accurately capture the actual execution process of the labeled actions, providing a high-quality data foundation for subsequent trajectory analysis, action recognition, and skill learning. This method has advantages such as strong adaptability, good robustness, and no need for manual intervention, making it suitable for trajectory segmentation requirements of various robotic arm operations.

Claims

1. A mechanical arm trajectory segmentation method based on Z-axis state conversion and speed trough, collecting the complete motion trajectory of the end effector of the mechanical arm in the Cartesian space, recording the timestamp sequence and the corresponding position sequence , extracting the Z-axis height sequence therefrom, and performing trajectory segmentation, characterized in that: characteristics of the Z-axis height sequence, set the relative height coefficient , ∈(0,1), calculate the relative low threshold value and the relative high threshold value ; wherein, , represent the global minimum and global maximum of the Z-axis height, ; According to the size relationship between the Z-axis height and the relative threshold value, a high state, an intermediate state and a low state are divided; a moment when the marker converts from the high state or the intermediate state to the low state is marked as a falling conversion point, and a moment when the marker converts from the low state to the high state or the intermediate state is marked as a rising conversion point; for each falling conversion point, a first rising conversion point is searched in a subsequent frame to pair with the falling conversion point, and the falling conversion point and the rising conversion point form a candidate interval; a candidate interval with the longest time span is selected from all candidate intervals as a rough candidate interval ; Calculating a speed magnitude sequence based on position coordinates of adjacent frames Performing speed valley detection on the speed magnitude sequence within the rough candidate interval to find the minimum value points of speed closest to the start and end points of the interval as the starting boundary candidate point and the ending boundary candidate point, constructing a boundary point pair based on the speed valley, segmenting the mechanical arm trajectory to obtain the action trajectory of task execution. ​ 2. The trajectory segmentation method for a robot arm based on Z-axis state transition and velocity trough according to claim 1, wherein: Using the 3σ criterion, the acquired original trajectory position sequence Data points deviating more than three standard deviations from the mean are considered outliers and removed; Gaussian filter parameters are set. The position sequence after outlier removal is smoothed and filtered, and then the Z-axis height sequence is extracted. 3.The trajectory segmentation method for a robot arm based on Z-axis state transition and velocity valley according to claim 1, wherein: Setting a relative height coefficient , .

4. The trajectory segmentation method for a robot arm based on Z-axis state transition and velocity trough according to claim 1, wherein: Based on a relatively low bit threshold and a relatively high bit threshold , the elements in the Z-axis height sequence are state-divided to form a state sequence : wherein, represents the first Z-axis coordinate component of the robot end effector in the Cartesian coordinate system of the frame, , represents the total number of frames of the trajectory.

5. The trajectory segmentation method for a robot arm based on Z-axis state transition and velocity trough according to claim 4, wherein: defining a set of falling transition points and a set of rising transition points respectively: For the set of falling transition points The falling transition point The rising transition point paired with it Is: wherein min denotes taking the minimum value; denotes a falling transition point corresponding timestamp, , denotes a set of rising transition points the rising transition point and its corresponding timestamp.

6. The trajectory segmentation method for a robot arm based on Z-axis state transition and velocity trough according to any one of claims 1-5, wherein: Computing time spans of candidate intervals , setting a minimum time span threshold and a maximum time span threshold , retaining only candidate intervals satisfying as valid pairing intervals, selecting an interval with the longest time span from the valid pairing intervals as a rough candidate interval .

7. The method of claim 6, wherein: If the set of falling transition points or the set of rising transition points is empty, or the time span of all candidate intervals does not satisfy , use the relatively low bit threshold to mark all frames that satisfy , find the longest continuous low bit state sequence as the action interval, and segment the robot arm trajectory.

8. The method of claim 1, wherein: The forward difference method is used to calculate the velocity vector sequence A Gaussian filter is used for smoothing, and a filter standard deviation is set .

9. The trajectory segmentation method for a robot arm based on Z-axis state transition and velocity trough according to any one of claims 1-8, wherein: define a front search window and a back search window , respectively, find the minimum velocity point within the front search window and the back search window , as the start boundary candidate point and the end boundary candidate point , construct the boundary point pair based on the velocity valley : wherein denotes the search window length.

10. The method of claim 9, wherein: For the boundary point pair , judge whether the following condition is satisfied , if not, let , carry out boundary constraint; use the interval after the boundary constraint as the final segmentation boundary , segment the mechanical arm trajectory to obtain the action trajectory of task execution.

Images