Arm trajectory planning method and control system for intelligent humanoid robot with body based on deep learning

Abstract

The invention discloses a deep learning-based arm trajectory planning method and control system for an intelligent humanoid robot with a body. The method comprises the following steps: acquiring a three-dimensional grid map of a multi-modal data construction environment; generating an initial feasible path through a path planning strategy network based on deep reinforcement learning; extracting the collision probability of each path node on the initial feasible path; according to the collision probability of each path node, judging an area in which local re-planning needs to be triggered; and carrying out smoothing processing to obtain a smooth obstacle avoidance path. The initial feasible path is generated through the path planning strategy network based on deep reinforcement learning. According to the collision probability of each node, an area needing local re-planning is screened out, and the path of the area is adjusted through an artificial potential field method, so that the capability of avoiding dynamic obstacles in the path planning process is improved, the collision risk in the path planning is reduced, and the path planning efficiency is improved. And the self-adaptive capability of the arm path planning complex environment of the intelligent humanoid robot with the body is improved.

Description

Technical Field

[0001] The present invention relates to the field of robot control technology, and in particular to a deep learning-based embodied intelligent humanoid robot arm trajectory planning method and control system. Background Art

[0002] With the increasing demand for intelligent humanoid robots, robotic arm path planning technology faces multiple challenges in dynamic, unstructured environments. Traditional path planning methods primarily assume deterministic environments, with typical examples including the graph-search-based A algorithm (Dijkstra et al., 1959), the random sampling-based RRT algorithm (LaValle et al., 1998), and its improved variants (such as RRT and Informed-RRT). These methods face the following technical bottlenecks in practical applications: Relying on a pre-built, high-precision environmental model (typically requiring obstacle position errors of ≤2cm), the entire planning process must be re-executed when encountering moving obstacles or sudden environmental changes. Experiments have shown that when obstacles move faster than 0.5m / s, traditional methods experience response delays of up to 300-500ms, failing to meet real-time obstacle avoidance requirements. For humanoid manipulators with 7 or more degrees of freedom (DoF), the increased spatial dimensionality of their configurations leads to an exponential increase in path search time. In MATLAB simulations, the RRT* algorithm achieves an average convergence time of 8.7 seconds in 6-DoF scenarios and 41.3 seconds in 7-DoF scenarios (approximately 4.7 times the time for each order of dimensionality increase). Key parameters in obstacle avoidance strategies, such as the safety distance threshold (typically set to 5-15cm) and path smoothness parameters (such as the order of the B-spline curve), require empirical adjustment through extensive trial and error. Industrial testing has shown that when laboratory parameter adjustment results are directly transferred to actual production lines, the collision probability increases by an average of 23.6% (comparative data for ABB YuMi robotic arms in electronic component assembly scenarios). Traditional methods fail to establish a mapping relationship between environmental disturbances and path corrections. When the end of the robotic arm is disturbed by external forces (such as a sudden 5 N·m torque shock), the current task must be completely interrupted and the path replanned, resulting in a task interruption rate as high as 62%.

[0003] In recent years, deep reinforcement learning (DRL) technology has provided new solutions for path planning. For example, the DDPG algorithm (Lillicrap, TP, 2015), the PPO algorithm (Schulman, J., 2017), and the GAIL solution (Ho, J., 2016), based on the actor-critic framework, have demonstrated environmental perception capabilities. However, existing deep learning solutions still have the following technical drawbacks: 1. Insufficient multimodal perception fusion: Existing models often rely on single visual input (RGB / RGB-D data), failing to effectively integrate heterogeneous information from multiple sources, such as force sensors (sampling rate ≥1kHz for 6-axis torque data) and inertial measurement units (IMU angular velocity accuracy 0.01° / s). On the MIT-Manus testbed, the end-point positioning error of a purely visual solution can reach ±3.2cm, while integrating force data reduces this error to ±0.8cm (experimental data released at ICRA2022).

[0004] 2. Lack of kinematic constraints: The existing network architecture does not explicitly encode physical constraints such as the robot arm's joint angle limits (such as ±150° rotation constraints) and link interference (collision detection based on DH parameters). Gazebo simulations show that approximately 17.4% of the paths generated by DRL have kinematic infeasibility issues (such as joint limit violations that cause servo motor overload).

[0005] 3. Inefficient online learning: Traditional DRL frameworks require millions of interaction samples for offline training (an average of 3.2 × 10^6 training steps in the MuJoCo environment). When parameter drift occurs in the deployment environment (such as a ±15% change in payload mass or a ±30% fluctuation in the friction coefficient), model adaptive adjustment takes more than 20 minutes (measured on the NVIDIA Jetson AGX platform).

[0006] 4. Insufficient energy optimization: Existing solutions fail to incorporate joint torque energy consumption into reward function design, resulting in the energy consumption of the generated path being an average of 38.7% higher than the optimal solution (comparison data from the KUKALBRiiwa robotic arm in an 8-hour continuous handling task).

[0007] The above technical defects result in the current system's performance being limited in the following typical scenarios: Industrial flexible assembly: When the workpiece position is randomly offset by ±10cm, the success rate of the traditional visual servo system drops to 54%.

[0008] Home service scenario: When faced with sudden human movement (speed > 1m / s), the response delay of the existing obstacle avoidance algorithm causes the safety distance violation rate to increase to 41%.

[0009] Medical rehabilitation applications: In the event of a sudden change in the patient's muscle tension (surface electromyography signal amplitude change > 200μV), a pure position control solution may cause excessive contact force (> 20N), posing a safety hazard.

[0010] These limitations severely restrict the practical application of humanoid robot arms in complex scenarios, and there is an urgent need to develop new path planning methods to break through the existing technical bottlenecks. Summary of the Invention

[0011] The present invention aims to provide a deep learning-based trajectory planning method and control system for an embodied intelligent humanoid robot arm, addressing existing issues such as poor adaptability, high computational complexity, and insufficient obstacle avoidance in dynamic environments. By integrating deep reinforcement learning (DRL) with 3D environmental perception technology, the present invention utilizes multimodal sensors to collect real-time environmental data, constructs a dynamic 3D environmental model, and integrates deep learning models to achieve efficient and accurate path planning.

[0012] In a first aspect, the present invention provides a deep learning-based method for embodied intelligent humanoid robot arm trajectory planning, comprising: Multimodal data including environmental point cloud data is collected through multimodal sensors.

[0013] A three-dimensional grid map of the environment is constructed based on multimodal data, and moving obstacles are introduced to generate a dynamic three-dimensional environment model.

[0014] The initial feasible path is generated by a path planning policy network based on deep reinforcement learning.

[0015] Extract the collision probability of each node on the initial feasible path. Based on the collision probability of each node, determine the area that needs to trigger local replanning. The condition for triggering local replanning is that the collision probability of n consecutive nodes exceeds the trigger threshold. n is the preset trigger number, which must be greater than or equal to 2.

[0016] The path obtained after local replanning is smoothed to obtain a smooth obstacle avoidance path.

[0017] Control the embodied intelligent humanoid robot arm to move along a smooth obstacle-avoiding path.

[0018] Preferably, the collision probability of the path node is obtained by a lightweight neural network whose input is the point cloud data within a preset range centered on the current path node and the segments of the current path within the corresponding range.

[0019] Preferably, the collision probability of the path node is expressed as: in, is the sigmoid function; is the distance between the current path node and the closest obstacle; is the collision detection threshold.

[0020] Preferably, the trigger number n is set to a value of 3-5.

[0021] Preferably, the local replanning process is to adjust the path of the target area by artificial potential field method. The objective function for determining the path is as follows: in, Is a path node i The distance to the nearest obstacle; Is a path node i The collision probability of 、 Distance , collision probability The weight of .

[0022] Preferably, the smoothing process adopts cubic uniform B-spline interpolation and is subject to constraints, including joint angular acceleration ≤ 1.5 rad / s² and end effector speed ≤ 0.8 m / s.

[0023] Preferably, the path planning strategy network includes an environmental feature extraction module, a timing modeling module and a strategy generation module. The environmental feature extraction module adopts a 3D convolutional neural network structure to process three-dimensional point cloud data and extract spatial features. The timing modeling module adopts a bidirectional long short-term memory network to fuse the historical motion trajectory with the spatial features output by the environmental feature extraction module, and output the timing context features. The strategy generation module receives the timing context features output by the timing modeling module, obtains the joint angular velocity and end effector posture at each moment based on proximal strategy optimization, and splices them to obtain the initial feasible path.

[0024] Preferably, the multimodal sensor includes an RGB-D camera, an inertial measurement unit installed on the end effector, and force sensors and joint encoders corresponding to each joint.

[0025] Preferably, the three-dimensional grid map includes the locations and motion trajectories of dynamic obstacles. The process of creating the three-dimensional grid map is as follows: first, point cloud data is fused with IMU pose data collected by an inertial measurement unit (IMU) using a tightly coupled SLAM algorithm to generate a static three-dimensional grid map. Next, a Kalman filter is used to predict the motion trajectories of dynamic obstacles.

[0026] In a second aspect, the present invention provides an embodied intelligent humanoid robot arm path control system, which is used to execute the aforementioned embodied intelligent humanoid robot arm path planning method; the embodied intelligent humanoid robot arm path control system includes a multimodal sensor, a map construction module, a path generation module and a path optimization module. The multimodal sensor is used to collect point cloud data, end effector posture, torque and angle of each joint. The map construction module is used to establish a three-dimensional grid map based on the multimodal data collected by the multimodal sensor. The path generation module is used to construct a path planning strategy network and generate an initial feasible path. The path optimization module is used to perform local replanning and smoothing on the initial feasible path to obtain a smooth obstacle avoidance path.

[0027] In a third aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; and the processor executes the aforementioned method for embodied intelligent humanoid robot arm trajectory planning.

[0028] In a fourth aspect, the present invention provides a readable storage medium storing a computer program; when the computer program is executed by a processor, it is used to implement the aforementioned embodied intelligent humanoid robot arm trajectory planning method.

[0029] The present invention has the following beneficial effects: This invention generates an initial feasible path through a path planning strategy network based on deep reinforcement learning. It then screens areas requiring local replanning based on the collision probability of each node and adjusts the path in these areas using an artificial potential field method. This improves the ability to avoid obstacles during path planning, reduces the risk of collisions during path planning, and enhances the adaptability of the embodied intelligent humanoid robot arm's path planning in complex environments.

[0030] The present invention introduces a bidirectional long short-term memory network into the path planning strategy network for generating the initial feasible path, combines the annular space characteristics and the historical motion trajectory to obtain the temporal context characteristics, improves the excellence of the initial feasible path, and also saves a large amount of computing resources for the subsequent local optimization and execution feedback links, thereby improving the efficiency and robustness of path planning as a whole.

[0031] The present invention realizes real-time perception and response to dynamic environments through multimodal sensor data fusion and deep reinforcement learning technology, and improves the efficiency of path planning calculation by optimizing deep reinforcement learning algorithms (such as PPO) to adapt to the real-time path planning requirements of high-degree-of-freedom robotic arms.

[0032] The present invention ensures the accuracy and continuity of path planning through a path smoothing optimization algorithm (such as B-spline curve interpolation method), thereby improving the execution efficiency and safety of the robot arm. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the structure of a humanoid robot arm performing path planning in an embodiment of the present invention.

[0034] Figure 2 A flow chart of constructing a three-dimensional grid map and generating a dynamic three-dimensional environment model in an embodiment of the present invention.

[0035] Figure 3 This is a workflow diagram of the path planning strategy network in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The present invention will be further described below with reference to the accompanying drawings.

[0037] Example A deep learning-based trajectory planning method for an embodied intelligent humanoid robot arm is used for path planning and control of the embodied intelligent humanoid robot arm.

[0038] like Figure 1 As shown, the embodied intelligent humanoid robot arm includes a plurality of articulated arms 1 that are connected in rotation in sequence. Each adjacent articulated arm 1 is driven to rotate relative to each other by a servo motor. A joint encoder is installed between each adjacent articulated arm 1. The joint encoder is used to detect the rotation angle of each joint. The end articulated arm 1 is installed with an end effector (not shown in the figure) and an end detection mounting structure 2. The end detection mounting structure 2 is installed with an inertial measurement unit IMU and a force sensor. The inertial measurement unit IMU and the force sensor are used to detect the spatial position and force magnitude of the end effector.

[0039] The embodied intelligent humanoid robot arm path planning method comprises the following steps: Step 1: Multimodal data acquisition and dynamic 3D modeling.

[0040] Multimodal sensors collect real-time data on the robot arm's joint states and surroundings to construct a dynamic three-dimensional environmental model. The multimodal sensors include an RGB-D camera, an inertial measurement unit (IMU), force sensors, and joint encoders. The RGB-D camera is mounted on the main body of the humanoid robot. The dynamic three-dimensional environmental model is constructed by fusing point cloud data acquired by the RGB-D camera with pose information from the IMU to generate a high-precision three-dimensional grid map (implemented in step 2). Force sensors are used to directly measure forces and torques during manipulation. Joint encoders measure the angles of each joint in real time (and angular velocity through differentiation). The position and velocity of dynamic obstacles are updated in real time using a Kalman filter algorithm.

[0041] Sensor system configuration: The camera uses an RGB-D camera (Intel RealSense D435i) and must be fixed in a stable position outside the robot's workspace to ensure an unobstructed global field of view. Its parameters are: resolution 1920×1080@30Hz, depth accuracy ±2%@1m (at room temperature), and a maximum effective detection range of 2m (error ≤±3%@1m, error ≤±8% at 3m). Note: By integrating IMU pose data with tightly coupled SLAM, the absolute positioning error can be reduced to ±0.8mm (for a workspace ≤1.5m).

[0042] The force sensor is a 6-axis force sensor (OnRobotHEX-E) and must be integrated into the robot arm's end effector. Its parameters are: range ±200N, sampling rate 1kHz, and noise density ≤0.05N / √Hz.

[0043] IMU (MPU6050): Gyroscope range: ±2000° / s, gyroscope noise density: 0.005° / √Hz (data sheet nominal value), measured noise ≤ 0.015° / √Hz after temperature compensation (-20°C to 60°C).

[0044] In this step, point cloud data, inertial data, joint angles, and data from the six-axis force sensor constitute multimodal data. The inertial data and joint angles are measured by the IMU and joint encoders, respectively.

[0045] Step 2: Data preprocessing and feature extraction.

[0046] The multimodal data obtained in step 1 is preprocessed, including point cloud denoising, data normalization, time series alignment, construction of a 3D grid map, and extraction of key features. The key features include obstacle information and joint status in the 3D grid map.

[0047] Among them, the point cloud denoising uses the statistical outlier removal (SOR) algorithm, setting the number of neighborhood points to 50 and the standard deviation threshold to 1.0. The purpose is to solve the outliers generated by the RGB-D camera due to environmental interference, prevent noise points from being mistakenly identified as obstacles and causing path planning failure, and ensure the authenticity of the environment representation for the DRL strategy network in step 3; data normalization is to map the joint angle to [- , ], and the end-point pose coordinates are normalized to the interval [-1, 1]. This aims to map multi-source sensor data to the same numerical interval and eliminate magnitude differences. Time series alignment uses inertial data as a benchmark, achieving temporal synchronization of multimodal data through linear interpolation. The sampling frequency is unified at 100Hz to compensate for the acquisition latency between the camera (30ms) and encoder (2ms). This ensures that the joint pose and point cloud space match at the same moment, providing accurate temporal correlation for the LSTM in step three.

[0048] The process of point cloud denoising is as follows: using the statistical outlier removal (SOR) algorithm, setting the number of neighborhood points k = 50, the standard deviation multiplier α = 1.0, and removing points that exceed Outliers of the range (where is the global average distance, is the global standard deviation); voxel downsampling, the original point cloud (about 300,000 points) is downsampled to 50,000 points, with a voxel size of 5 mm³.

[0049] like Figure 2 As shown in the figure, after point cloud denoising is completed, the environment modeling is performed through the point cloud data to obtain a three-dimensional grid map and extract the position and motion trajectory of dynamic obstacles. The specific process is as follows: (1) Point cloud fusion: The RGB-D point cloud collected by the camera is fused with the IMU pose data through a tightly coupled SLAM algorithm to generate a static three-dimensional grid map (resolution 5 mm³).

[0050] (2) Dynamic obstacle update: A Kalman filter (with a state vector of [position, velocity]^T and observation noise covariance Q = diag([0.01, 0.01, 0.01, 0.1, 0.1])) is used to predict the state of dynamic obstacles, with an update frequency of 10 Hz. The dynamic obstacle state includes the trajectory of the dynamic obstacle. By adding moving dynamic obstacles to the static 3D grid map, a dynamic 3D environment model is obtained.

[0051] (3) Coordinate alignment: Align the sensor coordinate system to the robot base coordinate system through hand-eye calibration (Eye-to-Hand mode).

[0052] The process of data normalization is as follows: the 7 joint values Linear mapping to [- , ], get the standard joint value , the formula is: in, It is The joint values ​​of the joints, is the minimum joint value, is the maximum joint value.

[0053] Among them, the end pose coordinates (obtained from each joint angle) are normalized to [-1, 1], and the rotation quaternion (qw, qx, qy, qz) maintains unit length.

[0054] Time series alignment: Using IMU data (100Hz) as the benchmark, cubic spline interpolation is used to achieve time synchronization between RGB-D (30Hz) and force sensing (1kHz) data, with a maximum time deviation of <2ms.

[0055] Step 3: Design and train a deep reinforcement learning policy network.

[0056] like Figure 3 As shown, a path planning strategy network based on deep reinforcement learning is constructed, which includes an environment feature extraction module, a time series modeling module and a strategy generation module.

[0057] The environmental feature extraction module uses a 3D convolutional neural network (3D-CNN) to process three-dimensional point cloud data and extract spatial features.

[0058] The temporal modeling module adopts a bidirectional long short-term memory network (Bi-LSTM) to fuse the historical motion trajectory with the spatial features extracted by the environmental feature extraction module to output temporal context features.

[0059] The strategy generation module outputs joint control instructions of joint angular velocity and end effector posture based on the proximal policy optimization (PPO) algorithm and the temporal context features output by the temporal modeling module.

[0060] Set the network architecture parameters, where the input of the environmental feature extraction module is the 64×64×64 three-dimensional grid map obtained in step 2 (where 0 / 1 indicates the presence of obstacles). The environmental feature extraction module includes the first three-dimensional convolutional layer (Conv3D), the first activation function (ReLU), the maximum pooling layer (MaxPool3D), the second three-dimensional convolutional layer (Conv3D), the second activation function (ReLU), the third three-dimensional convolutional layer (Conv3D), and the global average pooling layer (GlobalAvgPool3D). Its specific structure is: Conv3D(32,kernel=5,stride=2)→ReLU→MaxPool3D(2)→Conv3D(64,kernel=3)→ReLU→Conv3D(128,kernel=3)→GlobalAvgPool3D.

[0061] The input to the bidirectional long short-term memory network is the robot's motion state over 10 historical frames. The robot's motion state includes the joint angles and the end effector's position. The bidirectional long short-term memory network has 256 hidden units and a time step length T of 10.

[0062] The inputs to the strategy generation module are the spatial features output by the environmental feature extraction module and the contextual feature vector from the temporal modeling module. The contextual feature vector combines the historical motion state sequence and the current environmental features, resulting in a 512-dimensional vector (the 256-dimensional output of the bidirectional LSTM hidden layer is concatenated to obtain the 512-dimensional vector). The fully connected layer network has a structure of 512→256→7+6. The temporal feature vector is first input into one or two fully connected layers, mapping it from 512 dimensions to 256 dimensions. Reluctant Unit (ReLU) activation function is used. The final fully connected layer maps the 256-dimensional vector to the action dimension, i.e., 7-dimensional joint angular velocity and 6-dimensional end-point pose increment, for a total of 13 dimensions.

[0063] The output of the strategy generation module is a time series. This time series includes the 7-dimensional joint angular velocity and 6-dimensional end-point pose increment corresponding to each moment. The data at each moment in the time series is concatenated to form the "joint space motion path" and the "Cartesian space end-point path," i.e., the initial feasible path.

[0064] Direct optimization of the above training strategy, without verification, could amplify simulation errors. Therefore, we constructed 20 dynamic scenarios (obstacle speeds 0.1-2.0 m / s) in Gazebo and fed the multimodal data obtained in step 1 to generate 100,000 training data sets. We also randomly perturbed the lighting, friction coefficient (μ = 0.1-0.6), and payload mass (±15%). After deployment on the physical robot, we used a "meta-learning (MAML)" framework, inputting the current state, historical action sequences, and reward feedback, and updating the network parameters every 100 steps. This series of designs aims to bridge the gap between simulation and reality, establishing a continuous learning pipeline between simulation and reality. This allows the robot to "continuously evolve on the fly," addressing the scenario migration failure problem of traditional methods due to model rigidity.

[0065] Step 4: Dynamic obstacle avoidance and path optimization.

[0066] The path is optimized through a dynamic obstacle avoidance mechanism, the collision prediction subnetwork is used to evaluate the path safety, and the path is smoothed in combination with the gradient descent method.

[0067] The dynamic obstacle avoidance mechanism calculates the collision probability of each node in the path through a collision prediction subnetwork. If the probability exceeds a threshold, local path replanning is triggered. An artificial potential field (APF) and rapidly expanding random tree (RRT*) hybrid algorithm are used to generate obstacle avoidance paths, and the paths are adjusted locally or globally to cope with real-time changes in dynamic obstacles.

[0068] First, the path planning policy network trained in step 3 performs forward reasoning given the starting state and target end position to generate an initial feasible path from the current state to the target position. This initial feasible path, which takes into account the robot's dynamic constraints and environmental characteristics, serves as the global planning main path input for optimization in this step. The deep reinforcement learning policy network uses the "current environmental obstacle information" (static obstacles in the environment) provided in steps 1 and 2 to predict collisions when generating the path. The collision prediction subnetwork takes as input a local point cloud (1m³ area) and a planned path segment. It uses a lightweight MobileNetV3 neural network (α=0.75) and outputs a collision probability p∈[0,1]. Replanning is triggered if the collision probability p for three consecutive nodes exceeds 0.7. Specifically, the collision prediction subnetwork takes the local 1m³ point cloud and the planned path segment as input and quickly outputs the collision probability p, enabling "early prediction" of path nodes. If the predicted collision probability p exceeds 0.7 for multiple consecutive frames (three frames in this example), local replanning is immediately triggered, eliminating the need for global recalculation.

[0069] In some other embodiments, instead of using a lightweight neural network to calculate the collision probability, the collision probability is calculated by the following expression: as follows: in, is the sigmoid function; Is the current path node The closest obstacle the distance between them; is the collision detection threshold.

[0070] The specific process of local replanning involves locally adjusting the initial feasible path to account for dynamic obstacles using the Artificial Potential Field (APF) method. The parameters for the APF method are: repulsive force gain η = 0.8, range d0 = 0.3m; attractive force gain k = 1.2, target attraction radius r = 0.5m.

[0071] In the artificial potential field method, the optimization goal is to minimize the weighted sum of path length and collision risk. The objective function is constructed as follows: in, Is a path node i The distance to the nearest obstacle; Is a path node i The collision probability of 、 Distance , collision probability The weight of .

[0072] After local replanning, the locally adjusted initial feasible path is smoothed using a cubic uniform B-spline interpolation method to obtain a final, directly executable, smooth obstacle avoidance path. The number of control points in this cubic uniform B-spline interpolation method is equal to the number of path points + 2, and the constraints are joint angular acceleration ≤ 1.5 rad / s² and terminal velocity ≤ 0.8 m / s, ensuring continuity of joint motion and compliance with mechanical constraints.

[0073] Step 5: Real-time control and feedback The purpose of this step is to issue path instructions and fine-tune the strategy based on real-time feedback to improve execution accuracy and optimize energy consumption.

[0074] The control cycle is 20ms (50Hz). The smooth path (a series of joint angular velocity commands) obtained in step 4 is sent through the ros_control interface of ROS. The motor current and actual posture are collected in real time, and the tracking error Δe is calculated.

[0075] In order to comprehensively evaluate the current execution effect, a loss function in the form of weighted mean square error is set (used to measure the degree of deviation between the actual execution effect and the expected target), with the terminal positioning error weight being 0.7 and the energy consumption weight being 0.3.

[0076] If Δe>2mm for 10 consecutive frames (the error exceeds the preset threshold of 2mm) or the loss function continues to increase, the policy network parameter update is triggered. The update frequency is 10Hz, and the fine-tuning step size is 0.1×the learning rate in the PPO algorithm each time to ensure that the update amplitude is small to ensure execution safety.

[0077] The feedback from this step will return to step three (strategy grid training), forming a closed loop of "execution → feedback → update".

[0078] Implementation effect verification, dynamic obstacle avoidance test: In the presence of five moving obstacles (0.5-1.2 m / s, random direction changes), the obstacle avoidance success rate reached 98.3%. On the NVIDIA Jetson AGX Xavier platform (32 GB RAM, GPU FP16 acceleration), the average replanning time was 15 ms (95% confidence interval [12 ms, 18 ms]). The test conditions were a 7-DoF robotic arm, the number of dynamic obstacles was ≤5, and the moving speed was ≤1.5 m / s.

[0079] This compares to the path accuracy of traditional RRT*, which has an average end-point error of 3.2mm and a maximum error of 8.7mm. Within a 1m×1m×1m workspace, the proposed system achieves end-point repeatability of 0.8mm (tested under ISO9283). Calibration using a laser tracker (FARO Vantage) verifies the maximum error is 1.5mm (end-point accuracy is achieved using an impedance-controlled closed-loop, a force sensor feedback frequency of 1kHz, and position error compensation using PID control).

[0080] Finally, energy consumption was analyzed. For a standard grasping task (2kg load, 0.5m travel), the conventional method consumed a total of 1420J, while the proposed method consumed only 892J (test equipment: KUKALBRiiwa, sampling rate 1kHz). Supplementary experiments revealed that when the load varied by ±15%, the energy consumption fluctuated by ±8.7% (see Table 1 for details). Table 1 Comparison of path planning performance in multiple scenarios

Claims

1. A deep learning-based trajectory planning method for an embodied intelligent humanoid robot arm, characterized in that: include: Collecting multimodal data including environmental point cloud data through multimodal sensors; Build a 3D grid map of the environment based on multimodal data, and introduce moving obstacles to generate a dynamic 3D environment model; Generate an initial feasible path through a path planning strategy network based on deep reinforcement learning; Extract the collision probability of each node on the initial feasible path; determine the area that needs to trigger local replanning based on the collision probability of each node; the condition for determining whether to trigger local replanning is: the collision probability of n consecutive nodes is greater than the trigger threshold; n is the number of triggers, which must be greater than or equal to 2; The path obtained after local replanning is smoothed using APF+B-spline to obtain a smooth obstacle avoidance path.

2. The embodied intelligent humanoid robot arm trajectory planning method according to claim 1, characterized in that: The collision probability of the path node is obtained through a lightweight neural network; the input of the lightweight neural network is the point cloud data within a preset range centered on the current path node and the fragments of the current path within the corresponding range.

3. The embodied intelligent humanoid robot arm trajectory planning method according to claim 1, characterized in that: The expression of the collision probability of the path node is: ； in, is the sigmoid function; is the distance between the current path node and the closest obstacle; is the collision detection threshold.

4. The embodied intelligent humanoid robot arm trajectory planning method according to claim 1, characterized in that: The trigger number n is set to a value of 3 to 5.

5. The embodied intelligent humanoid robot arm trajectory planning method according to claim 1, characterized in that: The local replanning process is as follows: adjusting the path of the target area by the artificial potential field method; determining the objective function of the path is as follows: in, Is a path node i The distance to the nearest obstacle; Is a path node i The collision probability of 、 Distance , collision probability The weight of .

6. The embodied intelligent humanoid robot arm trajectory planning method according to claim 1, characterized in that: The smoothing method uses cubic uniform B-spline curve interpolation and sets constraints; the constraints include joint angular acceleration ≤ 1.5 rad / s² and end effector speed ≤ 0.8 m / s.

7. The embodied intelligent humanoid robot arm trajectory planning method according to claim 1, characterized in that: The path planning strategy network includes an environmental feature extraction module, a temporal modeling module, and a strategy generation module. The environmental feature extraction module uses a 3D convolutional neural network structure to process three-dimensional point cloud data and extract spatial features. The temporal modeling module uses a bidirectional long short-term memory network to fuse historical motion trajectories with the spatial features output by the environmental feature extraction module to output temporal context features. The strategy generation module receives the temporal context features output by the temporal modeling module, obtains the joint angular velocity and end effector posture at each moment based on proximal strategy optimization, and splices them to obtain an initial feasible path.

8. The embodied intelligent humanoid robot arm trajectory planning method according to claim 1, characterized in that: The multimodal sensor includes an RGB-D camera, an inertial measurement unit installed on the end effector, and force sensors and joint encoders corresponding to each joint.

9. The embodied intelligent humanoid robot arm trajectory planning method according to claim 8, characterized in that: The three-dimensional grid map contains the positions and motion trajectories of dynamic obstacles. The process of establishing the three-dimensional grid map is as follows: first, the point cloud data is fused with the IMU pose data collected by the inertial measurement unit through a tightly coupled SLAM algorithm to generate a static three-dimensional grid map; then, the motion trajectory of the dynamic obstacles is predicted using a Kalman filter.

10. A path control system for an embodied intelligent humanoid robot arm, characterized by: Used to execute the embodied intelligent humanoid robot arm trajectory planning method as described in claim 1; the embodied intelligent humanoid robot arm path control system includes a multimodal sensor, a map construction module, a path generation module and a path optimization module; the multimodal sensor is used to collect point cloud data, end effector posture, torque and angle of each joint; the map construction module is used to establish a three-dimensional grid map based on the multimodal data collected by the multimodal sensor; the path generation module is used to construct a path planning strategy network and generate an initial feasible path; The path optimization module is used to perform local replanning and smoothing on the initial feasible path to obtain a smooth obstacle avoidance path.

Images