Human-robot collaborative scene-oriented embodied robot dynamic obstacle avoidance prediction system

Abstract

This invention discloses a dynamic obstacle avoidance prediction system for embodied robots in human-robot collaboration scenarios, belonging to the field of intelligent obstacle avoidance technology for robots. It includes a perception fusion module, a trajectory prediction module, a safety situation assessment module, and a decision-making and planning module. The system constructs an environmental model by fusing multi-source sensor data and predicts multiple potential future movement trajectories of the human body based on a historical trajectory database. By calculating the probability distribution of collision risks between the robot and dynamic obstacles, an adaptive safety boundary is dynamically generated. Finally, based on this risk map and dynamic boundary, the robot's task trajectory is replanned to generate a collision-free movement trajectory. This technology improves the accuracy of predicting human movement intentions and enables intelligent real-time adjustment of the safety boundary, thereby enhancing the safety and operational efficiency of human-robot collaboration.

Description

Technical Field

[0001] This invention belongs to the field of intelligent obstacle avoidance technology for robots, specifically a dynamic obstacle avoidance prediction system for embodied robots in human-robot collaboration scenarios. Background Technology

[0002] In scenarios requiring close human-machine collaboration, such as industrial manufacturing, warehousing and logistics, and rehabilitation medicine, ensuring the safe and efficient operation of robots in dynamic, unstructured environments is crucial. Existing dynamic obstacle avoidance technologies for robots largely rely on real-time perception of the current position of moving objects in the environment and linear extrapolation of their short-term future positions based on simple physical motion models. This predictive approach simplifies human movement to a purely physical process, failing to consider the intentional and pattern-based nature of human behavior. In complex work areas or where multiple path choices exist, such methods struggle to accurately determine whether a person is heading towards workstation A or B, leading to significant deviations between predicted trajectories and actual conditions. This forces robots to adopt overly conservative or ineffective obstacle avoidance maneuvers.

[0003] To ensure safety, existing systems typically set a static safety boundary for robots with a fixed size, based on worst-case assumptions. This boundary remains unchanged throughout the robot's movement, regardless of the actual movement of people in the vicinity or the level of collision risk. When a person approaches, whether they walk directly towards the robot or merely pass by, the robot reacts at the same safe distance. While this strategy ensures basic safety, it severely sacrifices the robot's operational efficiency and the smoothness of its collaboration. In scenarios with frequent human-robot interaction, a fixed safety boundary leads to stiff robot movements, frequent starts and stops, and an inability to respond flexibly and intelligently based on real-time risk assessments, thus hindering the naturalness of human-robot collaboration and productivity. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art;

[0005] To this end, the present invention proposes a dynamic obstacle avoidance prediction system for embodied robots in human-robot collaboration scenarios, comprising:

[0006] The perception fusion module is used to acquire the robot's current motion state and environmental perception data collected by multiple sensors, and to fuse the environmental perception data to generate a dynamic environmental scene model.

[0007] The trajectory prediction module, based on the dynamic environment scene model, identifies moving human targets in space, extracts real-time motion trajectory segments of the human targets, and performs pattern matching on the real-time motion trajectory segments based on the historical trajectory database to predict several potential future motion trajectories of the human targets within a predefined time window.

[0008] The safety situation assessment module is used to obtain the robot's current motion state and preset task trajectory, combine it with several potential future motion trajectories of the human target, calculate the collision risk probability distribution map of spatial interference between the robot and all dynamic obstacles within the predefined time window, and dynamically adjust the robot's preset safety interaction boundary according to the collision risk probability distribution map to generate a dynamic safety boundary.

[0009] The decision planning module replans the robot's preset task trajectory based on the collision risk probability distribution map and the dynamic safety boundary, generating at least one collision-free candidate trajectory containing temporal path points and velocity sequences.

[0010] Furthermore, in the perception fusion module, the environmental perception data is fused, specifically as follows:

[0011] It receives point cloud data from a 3D vision sensor, RGBD data from a 2D image sensor, and pose change data from an inertial measurement unit;

[0012] The point cloud data is downsampled and denoised, and key feature point clouds are extracted.

[0013] The RGBD data is subjected to target detection and semantic segmentation to identify human bodies, workbenches, tools and temporary obstacles in the image, and each identified object is assigned a semantic label and a two-dimensional bounding box.

[0014] The two-dimensional bounding box and the point cloud data are spatiotemporally aligned and coordinate transformed, and the semantic labels are associated with the three-dimensional point cloud to form a structured point cloud with semantic information.

[0015] By combining pose change data from the inertial measurement unit, the structured point cloud in the robot body coordinate system is transformed into the global world coordinate system, and data from multiple consecutive frames are fused to construct the dynamic environment scene model. The dynamic environment scene model includes the three-dimensional spatial distribution information and velocity vectors of all dynamic and static obstacles in the space, specifically including the three-dimensional size, real-time position, movement speed, semantic category, and existence probability of each obstacle.

[0016] Furthermore, in the trajectory prediction module, pattern matching is performed on the real-time motion trajectory segments based on a historical trajectory database to predict several potential future motion trajectories of the human target, specifically:

[0017] From the dynamic environment scene model, the sequence of obstacle motion states labeled with human semantic categories is extracted as the real-time motion trajectory segment;

[0018] The real-time motion trajectory segment is encoded in a preset feature space and converted into a trajectory feature vector;

[0019] The similarity between the trajectory feature vector and the typical trajectory pattern features stored in the historical trajectory database is calculated. The typical trajectory pattern features correspond to different human intentions and task habits.

[0020] Based on the similarity ranking results, the top few most similar typical trajectory patterns are selected as the basic patterns for prediction.

[0021] For each selected typical trajectory pattern, based on the real-time position and velocity of the current human target and combined with the historical statistical characteristics of the typical trajectory pattern, the Gaussian process regression method is used to deduce multiple predicted trajectories that satisfy human kinematic constraints.

[0022] Each predicted trajectory is assigned a confidence weight, which is determined by the matching similarity and prediction uncertainty. All predicted trajectories and their weights together constitute the several potential future motion trajectories.

[0023] Furthermore, the safety situation assessment module calculates the collision risk probability distribution map, specifically as follows:

[0024] Obtain the robot's current motion state, including position, velocity, acceleration, and geometric model;

[0025] Obtain the preset task trajectory that the robot needs to execute, the preset task trajectory consists of a series of path points containing timestamps;

[0026] Obtain all potential future motion trajectories and their confidence weights output by the trajectory prediction module;

[0027] Within a predefined time window, with fixed time intervals as discrete moments, for each discrete moment, the predicted position and geometric envelope of the robot on its preset task trajectory are calculated.

[0028] For each potential future motion trajectory of each human target, calculate the predicted position and geometric envelope of the human body at the discrete time.

[0029] Calculate the minimum spatial distance between the geometric envelopes of the robot and the human body at this moment;

[0030] Based on the minimum spatial distance, the direction of motion speed of both parties, and the uncertainty of trajectory prediction, the instantaneous collision probability for the future potential motion trajectory at the discrete moment is calculated using a predefined risk assessment function.

[0031] The instantaneous collision probabilities under each trajectory are weighted and averaged according to their confidence weights to obtain the aggregate collision probability of the human target calculated at the current discrete moment.

[0032] Within a predefined time window, the calculation process of the aggregated collision probability is repeated for all discrete moments and all dynamic obstacles in space to generate a temporal and spatial probability field, namely the collision risk probability distribution map.

[0033] Furthermore, in the security situation assessment module, the preset security interaction boundary of the robot is dynamically adjusted to generate a dynamic security boundary, specifically as follows:

[0034] A basic safety interaction boundary value is predefined, which is determined based on the robot's maximum braking distance and the human comfort distance;

[0035] Extract the aggregated collision probability value at each time point in the spatial location from the collision risk probability distribution map;

[0036] Establish a mapping function from collision probability to safety boundary scaling factor, wherein the higher the aggregate collision probability, the larger the scaling factor, and the more the safety boundary expands;

[0037] Based on the aggregated collision probability near the robot's predicted position at each predicted time, the dynamic safety boundary value at the corresponding time is calculated using the mapping function.

[0038] The dynamic safety boundary values ​​are integrated into the geometric model of the robot body as additional constraints to form a dynamic protective shell that changes with time and space.

[0039] The outer contour of the dynamic protective shell is defined as the dynamic safety boundary, which is used for subsequent trajectory planning.

[0040] Furthermore, in the decision planning module, the robot's preset task trajectory is replanned to generate at least one collision-free candidate trajectory, specifically as follows:

[0041] The initial trajectory search space of the robot is defined by using the start and end points of the preset task trajectory as constraints;

[0042] Static obstacles in the dynamic environment scene model are treated as hard constraints;

[0043] Using the dynamic safety boundary as a time-varying constraint, it is required that the geometric envelope of the robot on the planned trajectory does not intrude into the dynamic safety boundary at any time.

[0044] Using the collision risk probability distribution map as the optimization objective, the planned trajectory is required to traverse low collision probability areas.

[0045] Within the trajectory search space, a sampling-based stochastic motion planning algorithm is used to generate a large number of feasible trajectories that satisfy the constraints.

[0046] For each feasible trajectory, calculate its total length, smoothness, and the comprehensive cost of accumulated collision risk;

[0047] Several feasible trajectories with the lowest overall cost are selected as the collision-free candidate trajectories for output.

[0048] Furthermore, the system also includes a motion control module for:

[0049] The optimal trajectory is selected from the collision-free candidate trajectories as the execution trajectory, and the robot body is driven to move according to the execution trajectory. At the same time, during the movement, the robot body is fine-tuned in real time according to the environmental information updated by the perception fusion module.

[0050] Furthermore, in the motion control module, the optimal trajectory is selected from the collision-free candidate trajectories as the execution trajectory, specifically as follows:

[0051] Receive several collision-free candidate trajectories output by the decision planning module;

[0052] For each collision-free candidate trajectory, additional real-time evaluation metrics are calculated, including continuity with the current robot motion state, torque requirements of joint actuators, and estimates of energy consumption.

[0053] Combining the comprehensive cost value calculated by the decision planning module with the real-time evaluation index, a final priority score is calculated for each candidate trajectory through a multi-objective decision function;

[0054] The candidate trajectory with the highest final priority score and no collision is selected and determined as the execution trajectory.

[0055] The execution trajectory is parsed into a sequence of position, velocity, and acceleration commands for each joint of the robot over continuous time.

[0056] The instruction sequence is sent to the underlying driver of the robot body to start executing the movement.

[0057] Furthermore, the motion control module performs real-time fine-tuning based on the environmental information updated by the perception fusion module, specifically as follows:

[0058] During the robot's movement along the execution trajectory, it continuously receives updated dynamic environment scene models provided by the perception fusion module;

[0059] The robot's current actual motion state is compared with the expected state of the execution trajectory, and the state error is calculated.

[0060] By utilizing the updated dynamic environment scenario model, we can predict new dynamic obstacles or state changes of predicted obstacles that will appear in the next control cycle.

[0061] Based on the model predictive control framework, rolling optimization is performed on subsequent segments of the execution trajectory within a local time window;

[0062] The optimization objective is to minimize state error while maintaining the overall task objective, and to avoid interference with obstacles in the updated environmental information.

[0063] The local trajectory adjustment obtained from the rolling optimization is converted into a correction amount for the joint control command, which is then superimposed on the original control command to achieve real-time fine-tuning of the motion.

[0064] Furthermore, the system also includes a historical trajectory learning module, which is used for:

[0065] During robot operation, the trajectory data of the human target identified by the trajectory prediction module and the final interaction results are continuously recorded.

[0066] Cluster analysis of large amounts of recorded trajectory data can reveal new repetitive movement patterns or specific operator's personal habit patterns.

[0067] The newly discovered motion patterns are feature-extracted and encoded, and then added to the historical trajectory database.

[0068] Based on the interaction results, the confidence weights of pattern matching in the historical trajectory database are adaptively updated, enabling the prediction model to continuously optimize as interaction experience accumulates.

[0069] Compared with the prior art, the beneficial effects of the present invention are:

[0070] Trajectory prediction technology based on pattern matching from historical trajectory databases can identify behavioral patterns and intentions of individuals based on short-term motion fragments, generating multiple logically consistent potential future trajectories and probability assessments. This method overcomes the limitations of traditional linear motion extrapolation, extending the prediction basis from instantaneous physical states to long-term behavioral patterns, thus improving the accuracy and temporal foresight of human motion trajectory prediction. It provides robots with more reliable and earlier predictive information, enabling obstacle avoidance decisions to be based on assessments of possible future paths for individuals, rather than simply reactions to their current position.

[0071] By calculating the probability distribution of collision risks and dynamically adjusting the safety boundaries accordingly, the robot system's safety strategy shifts from a fixed threshold to adaptive adjustment. This technology adjusts the shape and size of the safety zone around the robot in real time based on the real-time assessment of the future interference risk. During high-risk periods and directions, the boundary expands to proactively avoid collisions; during low-risk periods and directions, the boundary contracts to maintain operational efficiency. This achieves a real-time dynamic trade-off between safety protection and task execution, avoiding overly conservative behavior caused by static safety boundaries, reducing unnecessary motion interruptions and delays, and making human-robot collaborative operations smoother. Attached Figure Description

[0072] Figure 1 This is a timing diagram of the embodied robot dynamic obstacle avoidance prediction system for human-robot collaboration scenarios described in this invention.

[0073] Figure 2 A flowchart for predicting the potential future motion trajectory of a human target;

[0074] Figure 3 Flowchart for generating dynamic security boundaries;

[0075] Figure 4 A collision-free candidate trajectory map for robot 3D simulation;

[0076] Figure 5 Obstacle avoidance map for robot-human collaborative planar route planning. Detailed Implementation

[0077] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0078] See Figure 1The perception fusion module is responsible for acquiring the robot's current motion state and simultaneously collecting environmental perception data from multiple sensors. This heterogeneous data is fused and processed to construct a dynamic environmental scene model containing information on both dynamic and static obstacles. Subsequently, the trajectory prediction module, based on this model, specifically identifies moving human targets in space and extracts short-term motion states to form real-time motion trajectory segments. By matching these segments with typical motion patterns stored in the historical trajectory database, this module can predict multiple possible motion trajectories of the human body within a future period. The safety situation assessment module simultaneously acquires the robot's current state and its preset task trajectory. Combining this with the predicted human motion trajectory, it calculates the probability of the robot interfering with all dynamic obstacles within a predefined time window, forming a collision risk probability distribution map. This module also dynamically adjusts the robot's safe distance from the outside world based on this risk map, generating a dynamic safety boundary that changes over time and space. Finally, the decision planning module uses the collision risk probability distribution map as an optimization reference and the dynamic safety boundary as a hard constraint to replan the robot's original preset task trajectory, generating one or more feasible trajectories with no collision risk in the temporal path points and velocity sequences, thereby completing the closed loop from environmental perception to safe action decision-making.

[0079] In one embodiment of the present invention, the perception fusion module acquires and fuses multi-source sensor data. The perception fusion module receives point cloud data from a 3D vision sensor, RGBD data from a 2D image sensor, and pose change data from an inertial measurement unit. It performs downsampling and noise reduction processing on the point cloud data, extracts key feature point clouds from the point cloud data, performs target detection and semantic segmentation on the RGBD data, identifies human bodies, workbenches, tools, and temporary obstacle objects in the RGBD images, and assigns semantic labels and 2D bounding boxes to each identified object. The 2D bounding boxes are spatiotemporally aligned and coordinate transformed with the processed point cloud data, thereby associating the semantic labels with the 3D point cloud to form a structured point cloud with semantic information. Combined with the pose change data from the inertial measurement unit, the structured point cloud located in the robot body coordinate system is transformed to the global world coordinate system. By fusing data from multiple consecutive frames, a dynamic environment scene model is finally constructed. The dynamic environment scene model contains the 3D spatial distribution information and velocity vectors of all dynamic and static obstacles in the space. The specific information includes the 3D size, real-time position, movement speed, semantic category, and existence probability of each obstacle.

[0080] In some embodiments, downsampling reduces the number of point cloud data points to balance computational load and detail preservation. Noise reduction filters out outliers caused by sensor noise or environmental interference. Key feature point cloud extraction is based on local geometric features such as curvature or normal direction. Target detection and semantic segmentation of RGBD data employ a convolutional neural network (CNN) model. The CNN model outputs the class probability and bounding box coordinates for each pixel, and semantic labels and 2D bounding boxes are parsed from the CNN model's output. In specific implementations, spatiotemporal alignment and coordinate transformation rely on the synchronization of calibration parameters and timestamps between sensors. Mapping the 2D bounding box to point cloud data involves a transformation from the camera coordinate system to the robot's body coordinate system. Then, pose change data is used to unify the point cloud data in the robot's body coordinate system to the global world coordinate system. The mathematical expression for coordinate transformation is: ;

[0081] in: This represents a point coordinate vector in the global world coordinate system. This represents the point coordinate vector in the robot's body coordinate system. This represents the rotation matrix from the robot's body coordinate system to the global world coordinate system. The translation vector and rotation matrix represent the distance from the robot's body coordinate system to the global world coordinate system. Translation vector The pose change data is provided by real-time measurement by the inertial measurement unit. Optionally, the point cloud data is downsampled using a voxel grid method, which divides the three-dimensional space into regular grids and retains representative points within each grid. Noise removal is performed using a statistical filter to remove points far from the neighborhood mean. The extraction of key feature points is achieved by calculating the curvature value of each point in the point cloud data and selecting points with curvature higher than a threshold.

[0082] It is understandable that the construction of the dynamic environment scene model is an iterative process. The dynamic environment scene model is updated each time a new sensor data frame is received. The obstacle velocity vector in the dynamic environment scene model is obtained by differentiating the position of the same obstacle in two consecutive frames of the dynamic environment scene model and dividing by the time interval. The probability of obstacle existence is calculated based on the confidence score of target detection and the clustering stability of the point cloud data. In some embodiments, the association between semantic labels and 3D point clouds is achieved by projecting 2D bounding boxes into 3D space and matching them with the point cloud data clustering results. Successfully matched point cloud clusters are assigned corresponding semantic labels. Each point in the structured point cloud contains 3D coordinates, semantic category, and membership information. Optionally, the pose change data of the inertial measurement unit and the visual sensor data are fused using an extended Kalman filter to improve pose estimation accuracy. The extended Kalman filter outputs a smooth rotation matrix. Translation vector Used for coordinate transformation. In specific implementation, the fusion of continuous multi-frame data adopts a sliding window approach. The sliding window approach retains the dynamic environment scene model data of the most recent few frames and calculates a weighted average to generate the current dynamic environment scene model. The dynamic environment scene model is stored in a 3D mesh or octree data structure for easy retrieval. It can be understood that the multi-source sensors include LiDAR, depth camera, and inertial measurement unit. LiDAR provides dense point cloud data, depth camera provides RGBD data, and inertial measurement unit provides pose change data. All sensor data are synchronized through hardware or software timestamp alignment to ensure temporal consistency.

[0083] See Figure 2 In one embodiment of the present invention, the trajectory prediction module predicts human trajectory based on a dynamic environment scene model. The trajectory prediction module extracts the obstacle motion state sequence labeled with human semantic category from the dynamic environment scene model, uses the obstacle motion state sequence as a real-time motion trajectory segment, encodes the real-time motion trajectory segment in a preset feature space, converts the real-time motion trajectory segment into a trajectory feature vector, and calculates the similarity between the trajectory feature vector and the typical trajectory pattern features stored in the historical trajectory database. The typical trajectory pattern features correspond to different human intentions and task habits. According to the similarity ranking results, the top few most similar typical trajectory patterns are selected as the basic prediction patterns. For each selected typical trajectory pattern, based on the real-time position and speed of the current human target and combined with the historical statistical features of the typical trajectory pattern, the Gaussian process regression method is used to deduce multiple predicted trajectories that meet human kinematic constraints. Each predicted trajectory is assigned a confidence weight, which is jointly determined by the matching similarity and the prediction uncertainty. All predicted trajectories and their weights together constitute several potential future motion trajectories of the human target.

[0084] In some embodiments, the real-time motion trajectory segment includes a sequence of three-dimensional position coordinates and velocity vectors of a human target over the most recent sampling periods. A pre-defined feature space encoding process compresses the sequence information into a fixed-dimensional trajectory feature vector. This trajectory feature vector includes the directionality, curvature features, average velocity, and relative position information relative to key landmarks in the environment. In a specific implementation, similarity calculation is performed by measuring the distance between the trajectory feature vector and the feature vectors of each typical trajectory pattern in the historical trajectory database. One distance metric is defined by the following formula: ;

[0085] in: This represents a scalar value indicating the dissimilarity between the trajectory feature vector and the feature vector of a typical trajectory pattern. This represents the current trajectory feature vector obtained from encoding real-time motion trajectory segments. This represents a feature vector of a typical trajectory pattern stored in a historical trajectory database, with the symbol... This represents the dot product operation of vectors. This represents the Euclidean norm of the vector. Optionally, the historical trajectory database organizes typical trajectory pattern features in a graph structure. Nodes in the graph structure represent typical trajectory pattern features, and edges represent the transition probabilities between different typical trajectory pattern features. Similarity calculation and ranking are performed on the graph structure to accelerate retrieval.

[0086] It is understandable that the Gaussian process regression method uses historical trajectory samples of selected typical trajectory patterns as training data to establish a probability distribution model of human position changes over time. Using the current real-time position and velocity as initial conditions, the probability distribution model outputs a probability distribution of future positions. Multiple predicted trajectories satisfying human kinematic constraints are then sampled from this probability distribution. In some embodiments, these kinematic constraints are manifested as the curvature continuity of the predicted trajectory and the physically feasible range of velocity and acceleration. The kernel function selection for Gaussian process regression considers the smoothness of human motion, and the calculation of the confidence weights incorporates the dissimilarity scalar values ​​derived from the formula. The reciprocal of the covariance matrix predicted by the Gaussian process regression is used. In specific implementation, the historical trajectory database is maintained through offline learning and online updates. During the offline learning phase, a large amount of human motion trajectory data in human-computer collaboration scenarios is collected for cluster analysis to extract typical trajectory pattern features. During the online update phase, the system's historical trajectory learning module supplements and optimizes the historical trajectory database based on new interaction data. Optionally, the number of most similar typical trajectory patterns selected is a fixed parameter or based on the current dissimilarity scalar value. The dynamic threshold is determined such that for each generated predicted trajectory, it includes not only a three-dimensional spatial position sequence, but also a corresponding velocity and orientation sequence.

[0087] It is understandable that the trajectory prediction module operates periodically. Each time it receives an updated dynamic environment scene model from the perception fusion module, it initiates a new round of prediction. Multiple potential future motion trajectories are output as a trajectory set, with each predicted trajectory in the set accompanied by a normalized confidence weight, the sum of which is 1. In specific implementation, feature space encoding employs an autoencoder neural network. This network compresses variable-length trajectory sequences into low-dimensional trajectory feature vectors and attempts to reconstruct the sequence at the decoding end to ensure the trajectory feature vectors contain sufficient information. Typical trajectory pattern feature vectors in the historical trajectory database are obtained by performing the same autoencoder encoding on the cluster center trajectories. Optionally, for real-time motion trajectory segments that fail to highly match any typical trajectory pattern features in the historical trajectory database, the trajectory prediction module uses a constant velocity or constant turning rate model based on physical dynamics as an alternative to generate a basic prediction pattern.

[0088] See Figure 3 In one embodiment of the present invention, the safety situation assessment module calculates a collision risk probability distribution map and generates a dynamic safety boundary. The safety situation assessment module acquires the robot's current motion state, which includes position, velocity, acceleration, and geometric size model. Simultaneously, the safety situation assessment module acquires a preset task trajectory consisting of a series of path points with timestamps. The safety situation assessment module also receives all future potential motion trajectories and their confidence weights output by the trajectory prediction module. Within a predefined time window, with fixed time intervals as discrete moments, for each discrete moment, the predicted position and geometric envelope of the robot on the preset task trajectory are calculated. The geometric envelope is obtained by expanding the robot's geometric size model based on the predicted position and posture. For each future potential motion trajectory of each human target, the predicted position and geometric envelope of the human body at the discrete moment are calculated. The minimum spatial distance between the robot's geometric envelope and the human body's geometric envelope at this moment is calculated. Based on the minimum spatial distance, the robot's motion velocity direction, the human body's motion velocity direction, and the uncertainty of trajectory prediction, the instantaneous collision probability for the future potential motion trajectory at the discrete moment is calculated using a predefined risk assessment function. The instantaneous collision probabilities under each potential future motion trajectory are weighted and averaged according to the confidence weights corresponding to the potential future motion trajectories to obtain the aggregated collision probability of the human target at the current discrete moment. Within a predefined time window, the process of calculating the aggregated collision probability is repeated for all discrete moments and all dynamic obstacles in space, generating a temporal and spatial probability field, which is the collision risk probability distribution map.

[0089] In some embodiments, the predefined risk assessment function uses the minimum spatial distance as the main input. When the minimum spatial distance is less than a set threshold, the instantaneous collision probability increases sharply. The relative angle between the robot's motion velocity direction and the human's motion velocity direction is also included in the calculation; the smaller the relative angle, the higher the instantaneous collision probability. The uncertainty of trajectory prediction is represented by the determinant of the posterior covariance matrix and used as a probability scaling factor. The safety situation assessment module predefines a basic safety interaction boundary value, which is determined based on the robot's maximum braking distance and the human's comfort distance. The robot's maximum braking distance is calculated based on the robot's current speed and maximum deceleration. From the collision risk probability distribution map, the aggregated collision probability value at each discrete moment near the robot's predicted position is extracted. A mapping function is established from the aggregated collision probability value to the safety boundary scaling factor. The mapping function satisfies a monotonic relationship where the higher the aggregated collision probability value, the larger the scaling factor and the more inflated the safety boundary. Based on the aggregated collision probability value near the robot's predicted position at each prediction moment, the dynamic safety boundary value at the corresponding moment is calculated through the mapping function. One specific form of the mapping function is as follows: ;

[0090] in: This represents the dimensionless value of the calculated safety boundary scaling factor. A positive coefficient representing the control of the maximum expansion amplitude. The positive coefficient representing the growth slope of the control function. This represents the aggregated collision probability value extracted from the collision risk probability distribution map. The dynamic safety boundary value is integrated into the robot's geometric model as an additional constraint, forming a dynamic protective shell that varies with time and space. This dynamic protective shell expands uniformly outwards from the robot's geometric envelope by the distance defined by the dynamic safety boundary value. The outer contour of the dynamic protective shell is defined as the dynamic safety boundary.

[0091] Optionally, the calculation of the robot's maximum braking distance considers the robot's dynamic limits, while the human comfort distance is set to a fixed value based on human-robot interaction standards or adaptively adjusted according to the operator's identity. In specific implementations, the minimum spatial distance is calculated using the GJK algorithm between two convex geometric envelopes. The geometric envelopes are typically approximated by combinations of basic shapes such as cylinders, spheres, or cuboids to balance computational accuracy and speed. It can be understood that the collision risk probability distribution map is stored and represented on a discrete spatiotemporal grid, with each grid cell recording an aggregated collision probability value. The temporal resolution of the spatiotemporal grid is consistent with the discrete time interval, while the spatial resolution is set according to the environmental size and accuracy requirements. In some embodiments, the weighted average processing assigns a smaller contribution to predicted trajectories with lower confidence weights. For static obstacles, their aggregated collision probability is obtained directly from a table based on the fixed distance between the robot and the static obstacle's geometric envelope; the collision risk of static obstacles does not change over time. Optionally, different basic safety interaction boundary values ​​and scaling factors can be set for different body parts of the robot in the dynamic safety boundary calculation. For example, the basic safety interaction boundary value of the robotic arm's end effector is greater than that of the robot's torso. In practice, the dynamic protective shell is generated in real time. Each planning cycle updates the shape and size of the dynamic protective shell based on the latest collision risk probability distribution map. The dynamic safety boundary serves as a time-varying constraint, provided to the decision-making and planning module. It is understood that the predefined time window typically covers a longer period than a single robot planning cycle to ensure long-term prediction of potential risks. The length of the time window matches the prediction duration of the trajectory prediction module.

[0092] In one embodiment of the present invention, the decision planning module replans a preset task trajectory. The module uses the start and end points of the preset task trajectory as constraints to define the robot's initial trajectory search space, which is a configuration space containing the robot's joint positions, velocities, and time dimensions. The decision planning module uses static obstacles in the dynamic environment scene model as hard constraints, requiring that the robot's geometric envelope on the planned trajectory does not spatially overlap with the geometric model of the static obstacles at any time. The module also uses a dynamic safety boundary as a time-varying constraint, requiring that the robot's geometric envelope on the planned trajectory does not intrude into the time-varying protective shell region described by the dynamic safety boundary at any time. Finally, the module uses a collision risk probability distribution map as an optimization objective, guiding the planned trajectory to traverse as many low-collision-probability regions as possible on the collision risk probability distribution map. Within the initial trajectory search space, the decision planning module uses a sampling-based stochastic motion planning algorithm to generate a large number of feasible trajectories that satisfy the constraints.

[0093] In some embodiments, the sampling-based stochastic motion planning algorithm employs a kinematic fast random tree algorithm. This algorithm randomly samples nodes in the robot configuration space and expands them by connecting them through a kinematic model, continuously performing collision detection during the expansion process. Collision detection targets not only static obstacle models but also the three-dimensional time-varying constraint body formed by the dynamic safety boundary at each time step. For each random sampling point or trajectory segment, the aggregated collision probability value corresponding to that spatiotemporal location needs to be obtained by querying the collision risk probability distribution map. Optionally, the stochastic motion planning algorithm can employ a bidirectional fast random tree algorithm or a probabilistic path graph algorithm. The probabilistic path graph algorithm pre-constructs a graph structure representing collision-free connectivity in the configuration space. In specific implementations, the planning process performs feasibility verification for each expanded trajectory path. Feasibility verification includes checking whether the joint positions, velocities, and accelerations are within mechanical limits and whether the trajectory satisfies the constraints of the dynamic safety boundary at all discrete times. For each path that successfully connects the start and end points and passes the feasibility verification, the decision planning module records it as a feasible trajectory.

[0094] The decision-making and planning module calculates the total length, smoothness, and cumulative collision risk of each feasible trajectory, forming a comprehensive cost value. The formula for calculating the comprehensive cost value is as follows: ;

[0095] in: The comprehensive cost of a feasible trajectory. This represents the total length scalar of the feasible trajectory in joint space or task space. The smoothness cost scalar representing the feasible trajectory is obtained by integrating over the trajectory's acceleration or jerk. The cumulative collision risk cost scalar represents the feasible trajectory, obtained by integrating the aggregated collision probability values ​​of each spatiotemporal unit it passes through along the trajectory. , and These represent the weighting coefficients for total length cost, smoothness cost, and cumulative collision risk cost, respectively. The decision-making and planning module selects the comprehensive cost. The lowest-ranking feasible trajectories are output as collision-free candidate trajectories. Refer to Table 1 for the parameter configuration of the motion planning algorithm used for sampling.

[0096] Table 1: Parameter configuration table for sampling-based stochastic motion planning algorithm.

[0097]

[0098] It is understandable that the geometric model of static obstacles originates from the dynamic environment scene model constructed by the perception fusion module and is represented in the form of a 3D mesh or convex polyhedron. Hard constraints are implemented by determining whether the robot's geometric envelope intersects with these static models during collision detection. The dynamic safety boundary, as a time-varying constraint, has constraints whose VALIDATION_RES may differ at each discrete verification time. Verification requires querying the 3D surface of the dynamic safety boundary at the corresponding time. In some embodiments, the cumulative collision risk cost... The calculation considers not only the magnitude of the aggregate collision probability value, but also the duration for which the robot is in a high-risk area; the longer the duration, the greater the accumulated collision risk cost. The higher the overall cost, the more feasible the trajectories. The number of feasible trajectories with the lowest overall cost can be set according to task requirements. For example, the three feasible trajectories with the lowest overall cost can be selected as collision-free candidate trajectories.

[0099] Optional, smoothness cost The specific calculation can be performed using the sum of squares of the trajectory joint accelerations or the sum of squares of the jerk, with the total length cost... The calculations are performed in joint space to reflect mechanical wear, with weighting coefficients. , and The decision-making and planning module determines the trajectory through offline tuning or online adaptive strategies. In specific implementations, if the stochastic motion planning algorithm fails to find any feasible trajectory within the set number of MAX_ITERATIONS iterations, the module adopts a strategy of relaxing some constraints. For example, it may allow the trajectory to briefly intrude into a low-risk dynamic safety boundary region at a very low speed, or extend the total allowed motion time. The collision-free candidate trajectory finally output by the decision-making and planning module contains a series of time-stamped path points and corresponding joint velocity sequences. The path point sequence and velocity sequence together define the robot's motion.

[0100] See Figure 4 The figure presents multiple sets of collision-free candidate trajectories output by the decision planning module. The simulation space is constructed using a three-dimensional coordinate system (X, Y, Z coordinates, unit: meters). Red solid circles mark the trajectory start points, blue pentagrams mark the trajectory end points, and gray and black triangular pyramids represent static obstacles in the dynamic environment scene model. Three different curves correspond to three collision-free candidate trajectories: the blue solid line is trajectory 1 (optimal), the green dashed line is trajectory 2 (second best), and the red dotted line is trajectory 3 (third best). These trajectories are generated by a sampling-based stochastic motion planning algorithm. Under the premise of satisfying the hard constraints of static obstacles and the time-varying constraints of dynamic safety boundaries, the optimization objective is to minimize the comprehensive cost value of total length, smoothness, and cumulative collision risk. As can be seen from the figure, the optimal trajectory 1 has the shortest overall path and the best smoothness. Its spatial direction always avoids the geometric envelope of static obstacles and traverses low-collision-probability areas as much as possible. Trajectories 2 and 3, due to their higher comprehensive cost value, serve as the second best and third best candidate trajectories, respectively, providing redundant choices. The visualization results validated the trajectory replanning capability of the decision planning module and provided an intuitive reference for the motion control module to select the execution trajectory.

[0101] In one embodiment of the present invention, the system includes a motion control module that selects the optimal trajectory from collision-free candidate trajectories as the execution trajectory. The motion control module receives several collision-free candidate trajectories output by the decision planning module. The motion control module calculates additional real-time evaluation indicators for each collision-free candidate trajectory. The real-time evaluation indicators include the continuity cost with the current robot motion state, the estimated torque demand of the joint actuators, and the estimated energy consumption. The motion control module combines the comprehensive cost calculated by the decision planning module with the real-time evaluation indicators and calculates a final priority score for each collision-free candidate trajectory through a multi-objective decision function. The motion control module selects the collision-free candidate trajectory with the highest final priority score as the execution trajectory. The motion control module parses the execution trajectory into a sequence of position commands, a sequence of velocity commands, and a sequence of acceleration commands for each joint of the robot in continuous time. The motion control module sends the sequence of position commands, the sequence of velocity commands, and the sequence of acceleration commands to the underlying actuators of the robot body to start the motion execution.

[0102] During the robot's movement along the execution trajectory, the motion control module continuously receives updated dynamic environment scene models provided by the perception fusion module. The motion control module compares the robot's current actual motion state with the expected state of the execution trajectory, calculates the state error between the robot's current actual motion state and the expected state of the execution trajectory, and uses the updated dynamic environment scene model to predict the state changes of newly appearing dynamic obstacles or predicted obstacles in the next control cycle. Based on the model predictive control framework, the motion control module performs rolling optimization on subsequent segments of the execution trajectory within a local time window. The optimization objective of rolling optimization is to minimize the state error while maintaining the overall task objective, and to avoid interference with obstacles in the updated dynamic environment scene model. The motion control module converts the local trajectory adjustment amount obtained from rolling optimization into the correction amount of the joint control commands. The motion control module superimposes the correction amount of the joint control commands onto the original joint control commands to achieve real-time fine-tuning of the motion. The system also includes a historical trajectory learning module. During robot operation, the historical trajectory learning module continuously records the human target trajectory data identified by the trajectory prediction module and the final interaction results of human-computer interaction. The historical trajectory learning module performs cluster analysis on the large amount of recorded human target trajectory data. Through cluster analysis, the historical trajectory learning module discovers new repetitive motion patterns or specific operator's personal habit patterns. The historical trajectory learning module extracts and encodes the features of the newly discovered motion patterns and adds the newly discovered motion patterns after feature extraction and encoding to the historical trajectory database. The historical trajectory learning module adaptively updates the confidence weight of pattern matching in the historical trajectory database based on the final interaction results of human-computer interaction.

[0103] In some embodiments, the multi-objective decision function is a weighted sum of the comprehensive cost value and various real-time evaluation indicators after normalization. A specific form of the multi-objective decision function is as follows: ;

[0104] in: Indicates the first The final priority score of each collision-free candidate trajectory. Indicates the first The comprehensive value of a collision-free candidate trajectory Indicates the first The continuity cost scalar between a collision-free candidate trajectory and the current robot motion state. Indicates the first The estimated peak torque demand of the joint actuator for each collision-free candidate trajectory. Indicates the first Estimated motion energy consumption of a collision-free candidate trajectory. , , and These are the corresponding positive weight coefficients. Optional, continuity cost. The torque demand estimate is obtained by calculating the norm of the difference between the starting point of the collision-free candidate trajectory and the robot's current joint position and velocity. The energy consumption estimate is obtained by calculation along a collision-free candidate trajectory based on the robot's inverse dynamics model. It is approximated by integrating the product of joint torque and joint velocity.

[0105] It is understood that state errors include the position and posture errors of the robot's end effector and the angle errors of each joint. The model predictive control framework solves a finite-time optimization problem in each control cycle. The constraints of the optimization problem include new obstacle avoidance constraints derived from the updated dynamic environment scene model and the robot's dynamic constraints. In some embodiments, the historical trajectory learning module uses an incremental clustering algorithm, such as the online K-means algorithm, to perform clustering analysis on human target trajectory data. Newly discovered motion patterns need to accumulate a certain number of samples before being added to the historical trajectory database to ensure the universality of the patterns. The adaptive update process adjusts the initial confidence weights of relevant typical trajectory pattern features in the historical trajectory database during matching based on whether a collision or emergency stop occurs in the interaction result. In a specific implementation, the correction amount of the joint control command is obtained by solving a quadratic programming problem. The quadratic programming problem aims to minimize the tracking error and control quantity change. The rolling optimization and command correction process is repeated in each control cycle to achieve continuous fine-tuning. Optionally, the final priority score... After calculation, the motion control module checks if the highest score exceeds an execution threshold. Only when the highest score exceeds the threshold will the corresponding collision-free candidate trajectory be determined as the execution trajectory; otherwise, the motion control module will request the decision planning module to replan or trigger an emergency stop. The historical trajectory learning module updates its operations during system idle periods or in background threads to avoid performance interference with the real-time control loop. The historical trajectory database uses version management to support rollback and comparative analysis.

[0106] See Figure 5This diagram illustrates the core execution effects of the system's decision-making, planning, and motion control. The global motion space is constructed using a Cartesian coordinate system, with X and Y coordinates representing the robot's position in a two-dimensional plane (unit: meters). Green dots represent the robot's starting point, red dots represent the task's ending point, brown filled blocks represent static obstacles, orange dot sequences represent the human's dynamic trajectory, and orange dots represent the human's current position. The gray dashed line represents the original task trajectory, which directly traverses the area covered by static obstacles and the human's dynamic trajectory, posing a clear collision risk. The system generates two collision-free candidate trajectories through the trajectory prediction module: the purple dashed line is candidate trajectory 1, and the green dashed line is candidate trajectory 2. Both avoid static obstacles, but candidate trajectory 1 has a higher degree of spatial overlap with the human's dynamic trajectory, resulting in a greater cumulative collision risk. The blue solid line represents the optimal execution trajectory (replanning) output by the system. This trajectory is generated by the decision planning module using a multi-objective optimization algorithm. It uses the start and end points of the original task trajectory as constraints, static obstacles as hard constraints, and dynamic safety boundaries as time-varying constraints. Simultaneously, it uses the collision risk probability distribution map as the optimization objective, selecting the feasible path with the lowest overall cost from the trajectory search space. From the execution results, the optimal execution trajectory, through smooth path adjustment, avoids all static obstacles while maintaining a safe distance from the human's dynamic trajectory, minimizing the costs of path length and motion smoothness. This demonstrates the system's real-time obstacle avoidance and trajectory optimization capabilities in dynamic environments.

[0107] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A dynamic obstacle avoidance and prediction system for embodied robots in human-robot collaboration scenarios, characterized in that, include: The perception fusion module is used to acquire the robot's current motion state and environmental perception data collected by multiple sensors, and to fuse the environmental perception data to generate a dynamic environmental scene model. The trajectory prediction module, based on the dynamic environment scene model, identifies moving human targets in space, extracts real-time motion trajectory segments of the human targets, and performs pattern matching on the real-time motion trajectory segments based on the historical trajectory database to predict several potential future motion trajectories of the human targets within a predefined time window. The safety situation assessment module is used to obtain the robot's current motion state and preset task trajectory, combine it with several potential future motion trajectories of the human target, calculate the collision risk probability distribution map of spatial interference between the robot and all dynamic obstacles within the predefined time window, and dynamically adjust the robot's preset safety interaction boundary according to the collision risk probability distribution map to generate a dynamic safety boundary. The decision planning module, based on the collision risk probability distribution map and the dynamic safety boundary, replans the robot's preset task trajectory to generate at least one collision-free candidate trajectory containing temporal path points and velocity sequences. The safety situation assessment module calculates the collision risk probability distribution map, specifically as follows: Obtain the robot's current motion state, including position, velocity, acceleration, and geometric model; Obtain the preset task trajectory that the robot needs to execute, the preset task trajectory consists of a series of path points containing timestamps; Obtain all potential future motion trajectories and their confidence weights output by the trajectory prediction module; Within a predefined time window, with fixed time intervals as discrete moments, for each discrete moment, the predicted position and geometric envelope of the robot on its preset task trajectory are calculated. For each potential future motion trajectory of each human target, calculate the predicted position and geometric envelope of the human body at the discrete time. Calculate the minimum spatial distance between the geometric envelopes of the robot and the human body at this moment; Based on the minimum spatial distance, the direction of motion speed of both parties, and the uncertainty of trajectory prediction, the instantaneous collision probability for the future potential motion trajectory at the discrete moment is calculated using a predefined risk assessment function. The instantaneous collision probabilities under each trajectory are weighted and averaged according to their confidence weights to obtain the aggregate collision probability of the human target calculated at the current discrete moment. Within a predefined time window, the calculation process of the aggregated collision probability is repeated for all discrete moments and all dynamic obstacles in space to generate a temporal and spatial probability field, namely the collision risk probability distribution map. In the aforementioned security situation assessment module, the robot's preset security interaction boundaries are dynamically adjusted to generate dynamic security boundaries, specifically as follows: A basic safety interaction boundary value is predefined, which is determined based on the robot's maximum braking distance and the human comfort distance; Extract the aggregated collision probability value at each time point in the spatial location from the collision risk probability distribution map; Establish a mapping function from collision probability to safety boundary scaling factor, wherein the higher the aggregate collision probability, the larger the scaling factor, and the more the safety boundary expands; Based on the aggregated collision probability near the robot's predicted position at each predicted time, the dynamic safety boundary value at the corresponding time is calculated using the mapping function. The dynamic safety boundary values ​​are integrated into the geometric model of the robot body as additional constraints to form a dynamic protective shell that changes with time and space. The outer contour of the dynamic protective shell is defined as the dynamic safety boundary, which is used for subsequent trajectory planning.

2. The embodied robot dynamic obstacle avoidance prediction system for human-robot collaborative scenarios according to claim 1, characterized in that, In the perception fusion module, the environmental perception data is fused, specifically as follows: It receives point cloud data from a 3D vision sensor, RGBD data from a 2D image sensor, and pose change data from an inertial measurement unit; The point cloud data is downsampled and denoised, and key feature point clouds are extracted. The RGBD data is subjected to target detection and semantic segmentation to identify human bodies, workbenches, tools and temporary obstacles in the image, and each identified object is assigned a semantic label and a two-dimensional bounding box. The two-dimensional bounding box and the point cloud data are spatiotemporally aligned and coordinate transformed, and the semantic labels are associated with the three-dimensional point cloud to form a structured point cloud with semantic information. By combining pose change data from the inertial measurement unit, the structured point cloud in the robot body coordinate system is transformed into the global world coordinate system, and data from multiple consecutive frames are fused to construct the dynamic environment scene model. The dynamic environment scene model includes the three-dimensional spatial distribution information and velocity vectors of all dynamic and static obstacles in the space, specifically including the three-dimensional size, real-time position, movement speed, semantic category, and existence probability of each obstacle.

3. The embodied robot dynamic obstacle avoidance prediction system for human-robot collaborative scenarios according to claim 2, characterized in that, In the trajectory prediction module, pattern matching is performed on the real-time motion trajectory segments based on a historical trajectory database to predict several potential future motion trajectories of the human target, specifically: From the dynamic environment scene model, the sequence of obstacle motion states labeled with human semantic categories is extracted as the real-time motion trajectory segment; The real-time motion trajectory segment is encoded in a preset feature space and converted into a trajectory feature vector; The similarity between the trajectory feature vector and the typical trajectory pattern features stored in the historical trajectory database is calculated. The typical trajectory pattern features correspond to different human intentions and task habits. Based on the similarity ranking results, the top few most similar typical trajectory patterns are selected as the basic patterns for prediction. For each selected typical trajectory pattern, based on the real-time position and velocity of the current human target and combined with the historical statistical characteristics of the typical trajectory pattern, the Gaussian process regression method is used to deduce multiple predicted trajectories that satisfy human kinematic constraints. Each predicted trajectory is assigned a confidence weight, which is determined by the matching similarity and prediction uncertainty. All predicted trajectories and their weights together constitute the several potential future motion trajectories.

4. The embodied robot dynamic obstacle avoidance prediction system for human-robot collaborative scenarios according to claim 3, characterized in that, In the decision planning module, the robot's preset task trajectory is replanned to generate at least one collision-free candidate trajectory, specifically as follows: The initial trajectory search space of the robot is defined by using the start and end points of the preset task trajectory as constraints; Static obstacles in the dynamic environment scene model are treated as hard constraints; Using the dynamic safety boundary as a time-varying constraint, it is required that the geometric envelope of the robot on the planned trajectory does not intrude into the dynamic safety boundary at any time. Using the collision risk probability distribution map as the optimization objective, the planned trajectory is required to traverse low collision probability areas. Within the trajectory search space, a sampling-based stochastic motion planning algorithm is used to generate a large number of feasible trajectories that satisfy the constraints. For each feasible trajectory, calculate its total length, smoothness, and the comprehensive cost of accumulated collision risk; Several feasible trajectories with the lowest overall cost are selected as the collision-free candidate trajectories for output.

5. The embodied robot dynamic obstacle avoidance prediction system for human-robot collaborative scenarios according to claim 4, characterized in that, The system also includes a motion control module for: The optimal trajectory is selected from the collision-free candidate trajectories as the execution trajectory, and the robot body is driven to move according to the execution trajectory. At the same time, during the movement, the robot body is fine-tuned in real time according to the environmental information updated by the perception fusion module.

6. The embodied robot dynamic obstacle avoidance prediction system for human-robot collaborative scenarios according to claim 5, characterized in that, In the motion control module, the optimal trajectory is selected from the collision-free candidate trajectories as the execution trajectory, specifically as follows: Receive several collision-free candidate trajectories output by the decision planning module; For each collision-free candidate trajectory, additional real-time evaluation metrics are calculated, including continuity with the current robot motion state, torque requirements of joint actuators, and estimates of energy consumption. Combining the comprehensive cost value calculated by the decision planning module with the real-time evaluation index, a final priority score is calculated for each candidate trajectory through a multi-objective decision function; The candidate trajectory with the highest final priority score and no collision is selected and determined as the execution trajectory. The execution trajectory is parsed into a sequence of position, velocity, and acceleration commands for each joint of the robot over continuous time. The instruction sequence is sent to the underlying driver of the robot body to start executing the movement.

7. The embodied robot dynamic obstacle avoidance prediction system for human-robot collaborative scenarios according to claim 6, characterized in that, In the motion control module, real-time fine-tuning is performed based on the environmental information updated by the perception fusion module, specifically as follows: During the robot's movement along the execution trajectory, it continuously receives updated dynamic environment scene models provided by the perception fusion module; The robot's current actual motion state is compared with the expected state of the execution trajectory, and the state error is calculated. By utilizing the updated dynamic environment scenario model, we can predict new dynamic obstacles or state changes of predicted obstacles that will appear in the next control cycle. Based on the model predictive control framework, rolling optimization is performed on subsequent segments of the execution trajectory within a local time window; The optimization objective is to minimize state error while maintaining the overall task objective, and to avoid interference with obstacles in the updated environmental information. The local trajectory adjustment obtained from the rolling optimization is converted into a correction amount for the joint control command, which is then superimposed on the original control command to achieve real-time fine-tuning of the motion.

8. The embodied robot dynamic obstacle avoidance prediction system for human-robot collaborative scenarios according to claim 1, characterized in that, It also includes a historical trajectory learning module, which is used for: During robot operation, the trajectory data of the human target identified by the trajectory prediction module and the final interaction results are continuously recorded. Cluster analysis of large amounts of recorded trajectory data can reveal new repetitive movement patterns or specific operator's personal habit patterns. The newly discovered motion patterns are feature-extracted and encoded, and then added to the historical trajectory database. Based on the interaction results, the confidence weights of pattern matching in the historical trajectory database are adaptively updated, enabling the prediction model to continuously optimize as interaction experience accumulates.

Images