Space non-cooperative target attitude recognition and capture method based on CNN-DDPG and double mechanical arms

Abstract

The invention relates to a spatial non-cooperative target attitude recognition and capture method based on CNN-DDPG and double mechanical arms. The method comprises the steps of collecting image data of a non-cooperative target; based on the image data, identifying a target attitude and a position parameter; based on the target posture and the position parameters, path planning is conducted, and an action instruction of the continuous control quantity of the double mechanical arms is obtained; and the double mechanical arms are driven to execute actions based on the action instruction, and meanwhile, the position states of the mechanical arms and the target are fed back, so that the double mechanical arms synchronously arrive at the grabbing point to capture the rotating target. Target attitude information can be acquired in real time, and an efficient and collision-free double-mechanical-arm cooperative trajectory can be generated, so that the capture success rate and precision are improved.

Description

Technical Field

[0001] This invention relates to the field of space robotics technology, and in particular to a method for spatial non-cooperative target posture recognition and capture based on CNN-DDPG and dual robotic arms. Background Technology

[0002] In the acquisition of non-cooperative targets in space, traditional attitude estimation and path planning methods typically rely on known or assumed target motion characteristics. However, for freely tumbling non-cooperative targets, traditional methods have many limitations. The dynamic behavior of targets is complex and difficult to predict; their attitude and position are often influenced by multiple factors, such as the non-uniformity of target surface reflectivity and complex geometry. This makes attitude estimation methods based on traditional geometry and feature matching (such as SIFT and SURF) perform poorly in practical applications, as these methods are highly sensitive to changes in illumination and viewing angle, and are easily affected by noise.

[0003] Existing methods, such as visual perception systems based on convolutional neural networks (CNNs), while improving the accuracy of pose estimation to some extent, still face several challenges. Especially when dealing with complex and dynamic space environments, improving the generalization and adaptability of models remains an unsolved problem. Although deep learning techniques have demonstrated powerful feature extraction capabilities in visual perception, these methods perform well on specific training datasets. In real-world space applications, the models' adaptability to unfamiliar targets or under varying lighting and viewing conditions needs further enhancement.

[0004] Furthermore, traditional path planning methods, such as those based on inverse kinematics or the Rapid Random Tree (RRT) algorithm, while capable of handling the motion of certain known targets, become significantly less effective against non-cooperative targets that are freely tumbling. Because the target's posture changes rapidly and unpredictably, traditional methods lack sufficient flexibility and real-time adaptability. In recent years, deep reinforcement learning (DRL) methods, especially the Deep Deterministic Policy Gradient (DDPG) algorithm, have been applied to space robot control, enabling dynamic adjustment of motion strategies through interaction with the environment to cope with the uncertain motion of the target.

[0005] In summary, existing technologies still suffer from insufficient accuracy, poor adaptability, and high computational complexity. To more effectively capture freely tumbling non-cooperative targets, a novel approach that tightly integrates perception and path planning is urgently needed to adapt to complex and dynamically changing space environments. Summary of the Invention

[0006] The purpose of this invention is to propose a spatial non-cooperative target pose recognition and capture method based on CNN-DDPG and dual robotic arms. The aim is to achieve high-precision pose estimation through convolutional neural networks (CNN) and combine deep reinforcement learning (DDPG) algorithm to perform synchronous and collision-free path planning for dual robotic arms, thereby improving the success rate and accuracy of spatial non-cooperative target capture.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A method for spatial non-cooperative target pose recognition and capture based on CNN-DDPG and dual robotic arms includes:

[0009] Acquire image data of non-cooperative targets;

[0010] Based on the image data, the target pose and position parameters are identified;

[0011] Based on the target posture and position parameters, path planning is performed to obtain motion commands for continuous control of the dual robotic arms.

[0012] Based on the action command, the dual robotic arms are driven to perform actions, and the position status of the robotic arms and the target is fed back, so that the dual robotic arms reach the grasping point synchronously and capture the rotating target.

[0013] Optionally, the method for identifying the target pose and position parameters is as follows: adopting a pose recognition model; the pose recognition model uses a ResNet101 residual network as the feature extraction network, and adds a fully connected regression layer at the end of the network.

[0014] Optionally, identifying target pose and position parameters includes:

[0015] The image data is preprocessed;

[0016] The preprocessed image data is input into the ResNet101 residual network, and computation is performed through residual blocks. To obtain deep features; where x is the input feature tensor, w is the residual block weight matrix, F(x,w) represents the nonlinear mapping inside the residual block, and y represents the output feature tensor;

[0017] The deep features are input into a fully connected regression layer to regress the target pose quaternion and relative position, thereby obtaining the target pose and position parameters.

[0018] Optionally, the preprocessing includes: normalizing and resizing the image to form an input tensor.

[0019] Optionally, path planning based on the target attitude and position parameters includes:

[0020] The target attitude and position parameters are used to form a state space as the current environment state vector, which is then input into a depth-deterministic policy gradient algorithm for path planning.

[0021] Optionally, the deep deterministic policy gradient algorithm includes: an Actor network and a Critic network; the Actor network, as the policy network, generates continuous control action commands for the robotic arm based on the state, and the Critic network evaluates the value of the output continuous control actions and uses them for policy updates and network training.

[0022] The method for implementing path planning using the deep deterministic policy gradient algorithm includes: ① constructing a state vector with the target pose / position and the states of both arms; ② outputting continuous control actions through the Actor network; ③ evaluating the value of the actions through the Critic network; ④ continuously optimizing through policy gradient; and ⑤ ultimately forming a cooperative path that can be updated in real time, is continuous, and is collision-free.

[0023] Optionally, after obtaining the motion commands for the continuous control quantities of the dual robotic arms, the process further includes:

[0024] The potential collision risk between the two robotic arms is detected by using a cylindrical envelope model. If a collision risk is detected, the motion commands are readjusted.

[0025] Optionally, driving the dual robotic arms to perform actions based on the action commands, while simultaneously feeding back the position status of the robotic arms relative to the target, includes:

[0026] After the dual robotic arms are driven to perform actions based on the aforementioned motion commands, the target posture and position parameters are re-identified, thereby dynamically correcting the trajectory and achieving closed-loop control.

[0027] The beneficial effects of this invention are as follows:

[0028] Compared with existing non-cooperative target acquisition technologies in space, the method proposed in this invention, based on CNN-DDPG and dual robotic arm collaborative control, can achieve high-precision real-time attitude recognition and safe and stable acquisition trajectory planning under the conditions of target free rolling and rapid attitude changes. It significantly improves the acquisition success rate and reduces the planning time, and has better applicability and robustness. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1This is a schematic diagram of a spatial non-cooperative target pose recognition and capture method based on CNN-DDPG and dual robotic arms according to an embodiment of the present invention;

[0031] Figure 2 This is a diagram of the attitude mapping network structure according to an embodiment of the present invention;

[0032] Figure 3 This is a structural diagram of the DDPG algorithm according to an embodiment of the present invention;

[0033] Figure 4 This is a sampled image of a non-cooperative target in space according to an embodiment of the present invention;

[0034] Figure 5 The training curves for the pose recognition network in this embodiment of the invention are shown.

[0035] Figure 6 The pose recognition network recognition error in this embodiment of the invention;

[0036] Figure 7 This is a DDPG training reward convergence graph according to an embodiment of the present invention;

[0037] Figure 8 This is a curve showing the change in distance between the left robotic arm and the capture point over time in an embodiment of the present invention.

[0038] Figure 9 This is a curve showing the change in distance between the right robotic arm and the capture point over time in an embodiment of the present invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0040] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] like Figure 1 As shown, this embodiment proposes a spatial non-cooperative target pose recognition and capture method based on CNN-DDPG and dual robotic arms, including:

[0042] Acquire image data of non-cooperative targets;

[0043] Based on the image data, the target pose and position parameters are identified;

[0044] Based on the target posture and position parameters, path planning is performed to obtain motion commands for continuous control of the dual robotic arms.

[0045] Based on the action command, the dual robotic arms are driven to perform actions, and the position status of the robotic arms and the target is fed back, so that the dual robotic arms reach the grasping point synchronously and capture the rotating target.

[0046] Specifically, to improve the autonomous recognition and capture performance of non-cooperative targets in space, this embodiment proposes an integrated intelligent attitude recognition and dual-arm path planning method based on a convolutional neural network (CNN) and a deep deterministic policy gradient (DDPG) algorithm. This method achieves real-time attitude perception and adaptive capture control of freely rotating targets by fusing deep learning-based visual recognition and reinforcement learning-based path planning. The system acquires target images using monocular vision, performs attitude recognition using a CNN model, and uses the recognition results as input to drive a DDPG-based dual-arm cooperative trajectory planner to generate continuous, collision-free, and synchronized capture paths, thereby improving the accuracy and robustness of capturing non-cooperative targets in space.

[0047] Furthermore, the method for identifying the target pose and position parameters is as follows: a pose recognition model is adopted; the pose recognition model uses a ResNet101 residual network as the feature extraction network, and a fully connected regression layer is added at the end of the network.

[0048] Furthermore, identifying target pose and position parameters includes:

[0049] The image data is preprocessed;

[0050] The preprocessed image data is input into the ResNet101 residual network, and computation is performed through residual blocks. To obtain deep features; where x is the input feature tensor, w is the residual block weight matrix, F(x,w) represents the nonlinear mapping inside the residual block, and y represents the output feature tensor;

[0051] The deep features are input into a fully connected regression layer to regress the target pose quaternion and relative position, thereby obtaining the target pose and position parameters.

[0052] Furthermore, based on the target pose and position parameters, path planning includes:

[0053] The target attitude and position parameters are used to form a state space as the current environment state vector, which is then input into a depth-deterministic policy gradient algorithm for path planning.

[0054] Furthermore, the deep deterministic policy gradient algorithm includes: an Actor network and a Critic network; the Actor network, as the policy network, generates continuous control action commands for the robotic arm based on the state, and the Critic network evaluates the value of the output continuous control actions and uses them for policy updates and network training.

[0055] The method for implementing path planning using the deep deterministic policy gradient algorithm includes: ① constructing a state vector with the target pose / position and the states of both arms; ② outputting continuous control actions through the Actor network; ③ evaluating the value of the actions through the Critic network; ④ continuously optimizing through policy gradient; and ⑤ ultimately forming a cooperative path that can be updated in real time, is continuous, and is collision-free.

[0056] Furthermore, after obtaining the motion commands for the continuous control quantities of the dual robotic arms, the following are also included:

[0057] The potential collision risk between the two robotic arms is detected by using a cylindrical envelope model. If a collision risk is detected, the motion commands are readjusted.

[0058] Furthermore, the action commands drive the dual robotic arms to perform actions, while simultaneously providing feedback on the position status of the robotic arms relative to the target, including:

[0059] After the dual robotic arms are driven to perform actions based on the aforementioned motion commands, the target posture and position parameters are re-identified, thereby dynamically correcting the trajectory and achieving closed-loop control.

[0060] This embodiment includes a visual perception unit, an intelligent path planning unit, and a closed-loop control execution module;

[0061] The visual perception unit is used to acquire monocular images and extract features using a convolutional neural network to achieve real-time recognition of target pose and position;

[0062] The intelligent path planning unit generates the cooperative motion trajectory of the two robotic arms based on the deep deterministic policy gradient (DDPG) algorithm and combined with the attitude information provided by the perception unit.

[0063] The closed-loop control execution module drives the dual robotic arms to move according to the planning results and feeds back the execution status to the sensing unit, realizing closed-loop optimization that integrates perception, decision-making and execution.

[0064] The working principle of this embodiment is as follows: When the servicing spacecraft performs a capture mission on a non-cooperative target, the system first acquires target images through the visual perception unit and extracts image features and target attitude using a CNN network. Subsequently, the identified attitude and position parameters are input into the deep reinforcement learning path planning module, where DDPG generates continuous control inputs for the dual robotic arms. The control execution unit drives the dual robotic arms to perform actions according to the planned instructions, while simultaneously providing feedback on the position status of the robotic arms and the target, achieving closed-loop optimization of perception-decision-control, enabling the robotic arms to complete cooperative capture even when the target is rotating at high speed.

[0065] The specific process of this embodiment is as follows:

[0066] (1) Target image acquisition and preprocessing: Real-time acquisition of continuous image frames of the freely tumbling target using an airborne monocular camera, followed by image normalization and resizing preprocessing to form the input tensor. To meet the input requirements of the neural network, where...

[0067] (2) Feature extraction using ResNet101 convolutional neural network: The processed image is input into the ResNet101 residual network, and features are extracted through residual blocks. It enables cross-layer transfer of deep features, achieves deep feature extraction, and enhances the ability to resist changes in illumination and attitude disturbances.

[0068] (3) The pose regression network outputs quaternions and position vectors: such as Figure 2 As shown, the high-dimensional feature vector is input into a fully connected regression network to regress the target pose quaternion. The relative positions are used, and the root mean square error (RMSE) is used as the loss function for training and optimization.

[0069] (4) Construct the state vector and input it into the DDPG module: Combine the target pose and position output by the CNN with the joint angles of the robotic arm to form the state space as the current environment state vector. Input the DDPG route planning module.

[0070] (5) DDPG control strategy generates dual-arm control inputs: The Actor network of DDPG outputs continuous control actions based on the state, and the Critic network evaluates the value of these actions and uses them for policy updates; the DDPG algorithm is as follows: Figure 3 As shown.

[0071] (6) Kinematics calculation and collision detection of robotic arms: Potential collision risks between the two robotic arms are detected using a cylindrical envelope model. If a collision risk is detected, the motion commands are readjusted.

[0072] (7) Closed-loop feedback execution and dynamic trajectory correction: The robotic arm executes the path movement and feeds back the execution status to the perception module in real time, so that the CNN can re-identify the target posture and dynamically correct the trajectory to achieve closed-loop control.

[0073] (8) Capture completion and stabilization control: The two robotic arms arrive at the gripping point simultaneously to achieve stable clamping of the rotating target and complete the capture operation.

[0074] To verify the effectiveness of the proposed CNN-based non-cooperative target pose recognition and DDPG-based dual-arm path planning methods in this embodiment, a complete simulation and training environment was constructed, including a visual data generation module, a pose recognition network training module, a reinforcement learning dual-arm planning and control module, and a capture execution verification module. Training images were generated by simulating a satellite 3D model, a deep neural network was trained to achieve real-time target pose estimation, and reinforcement learning was then used to achieve cooperative capture by the robotic arms.

[0075] (1) Training and accuracy verification of pose recognition network

[0076] This embodiment generates a monocular image dataset based on a 3D model of a non-cooperative target satellite to train a CNN pose recognition model. The observation distance is set to 20m, and the target is uniformly sampled within the range of [-180°, 180°] around three axes, generating one image every 5°, resulting in a total of 723 labeled image samples. The dataset is divided into an 80% training set and a 20% validation set, and the images undergo enhancement processing such as random rotation, brightness perturbation, and Gaussian noise to improve the network's generalization ability. A partial sample image of the non-cooperative target in space is shown below. Figure 4 As shown.

[0077] To train the intelligent pose recognition module of this invention, ResNet101 was used as the feature extraction network, and a fully connected regression layer was added at the end of the network to output quaternion pose parameters. During network training, the loss function decreased continuously with the number of iterations and tended to stabilize. Figure 5 As shown. Training results indicate that the average error in pose recognition is approximately 0.5°, and the position estimation error is controlled within 0.05 m.

[0078] During the dynamic testing phase, the initial angular velocity of the target was set to [0.2, 0.03, 0.05] rad / s. Under free nutation and high-speed rolling conditions, the attitude recognition network consistently output stable attitude estimates, with the attitude error remaining within ±0.5°. This verifies the real-time performance and robustness of this embodiment in dynamic target acquisition scenarios. Figure 6 As shown.

[0079] (2) Path planning training and collaborative control verification of dual robotic arms DDPG

[0080] The pose quaternion output by the pose recognition module in step (1) and the target position are used as the state input of the reinforcement learning environment. The DDPG deep deterministic policy gradient algorithm is used to train the collaborative path planning strategy of the two robotic arms. The training parameters are shown in Table 1 below:

[0081] Table 1 Training parameters

[0082] parameter set up Optimizer Adam Discount factor 0.9 Actor / Critic learning rate 0.002 Replay Buffer Size 120000 Number of training samples per batch 100 Actor / Critic Hidden Neural Layers 3 Number of hidden neurons per layer 256 Activation function RuLe / Tanh noise 0.1

[0083] The initial angular velocity of the target was set to [0.2, 0.03, 0.05] rad / s, and the capture points were set at both ends of the target's x-axis. The initial distance between the dual robotic arms and the target capture points was approximately 1.5 m. With training iterations, the reward function increased from approximately -80 to 42 and tended to converge, indicating that the strategy gradually reached its optimum. The DDPG training reward convergence is shown below. Figure 7 As shown.

[0084] With the CNN pose recognition and DDPG path planning modules operating in a closed loop, the dual robotic arms achieved synchronous capture, and the control effect is shown in Table 2 below:

[0085] Table 2 Control Effect

[0086] index result End-of-arm error ＜0.02m Time to capture and hold ≈18s Clamping synchronization error ＜0.003s

[0087] Figure 8 The curve showing the change in distance between the left robotic arm and the capture point over time is displayed; Figure 9 The curve showing the distance between the right robotic arm and the capture point as a function of time is displayed.

[0088] The results demonstrate that the method in this embodiment achieves high-precision attitude estimation (≈0.5°) and collision-free coordinated control, effectively overcoming the limitations of traditional geometry and model-dependent methods. It has good applicability and can effectively achieve efficient capture of targets that are freely tumbling in space under complex dynamic environments.

[0089] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A space non-cooperative target attitude identification and capture method based on CNN-DDPG and double robot arms, characterized in that, The method comprises the following steps: Collecting image data of a non-cooperative target; Identifying target attitude and position parameters based on the image data; Based on the target attitude and position parameters, path planning is performed to obtain the action instruction of the continuous control quantity of the dual-robot arm; Based on the action instruction, the dual-robot arm is driven to perform an action, and the state of the robot arm and the target position is fed back at the same time, so that the dual-robot arm synchronously reaches the grabbing point and captures the rotating target.

2. The CNN-DDPG and dual-arm based space non-cooperative target attitude identification and capture method according to claim 1, characterized in that, The method for identifying the target attitude and position parameters is to use an attitude recognition model; the attitude recognition model uses a ResNet101 residual network as a feature extraction network, and adds a fully connected regression layer at the end of the network.

3. The method of claim 2, wherein, The method for identifying the target attitude and position parameters comprises the following steps: Preprocessing the image data; The preprocessed image data is input into a ResNet101 residual network, and deep features are obtained by calculating a residual block ; wherein x is an input feature tensor, w is a residual block weight matrix, F(x, w) represents a nonlinear mapping inside the residual block, and y represents an output feature tensor. Inputting the deep features into the fully connected regression layer to regress the target attitude quaternion and relative position, and obtaining the target attitude and position parameters.

4. The method of claim 3, wherein, The preprocessing comprises normalizing and resizing the image to form an input tensor.

5. The method of claim 1, wherein, Based on the target attitude and position parameters, path planning comprises the following steps: The target attitude and position parameters are combined into a state space as a current environment state vector, which is input into a deep deterministic policy gradient algorithm to perform path planning.

6. The method of claim 5, wherein, The deep deterministic policy gradient algorithm comprises an Actor network and a Critic network; the Actor network serves as a policy network to generate continuous control action instructions for the robot arm according to the state, and the Critic network evaluates the value of the output continuous control action and is used for policy updating and network training. The method for implementing path planning by the deep deterministic policy gradient algorithm comprises the following steps: ① forming a state vector with the target attitude / position + dual-arm state, ② outputting continuous control actions by the Actor network, ③ evaluating the value of the actions by the Critic network, ④ continuously optimizing through policy gradient, and ⑤ finally forming a real-time updateable, continuous, and collision-free cooperative path.

7. The method of claim 1, wherein the CNN-DDPG and dual-arm space non-cooperative target pose identification and capture method is characterized in that, After obtaining the action instruction of the continuous control quantity of the dual-robot arm, the following steps are further included: Detecting potential collision risks between the dual-robot arms through a cylindrical envelope model, and readjusting the action instruction if a collision risk is found.

8. The method of claim 1, wherein the CNN-DDPG and dual-arm space non-cooperative target pose identification and capture method is characterized in that, Based on the action instruction, the dual-robot arm is driven to perform an action, and the state of the robot arm and the target position is fed back at the same time, so that the dual-robot arm synchronously reaches the grabbing point and captures the rotating target. After driving the dual-robot arm to perform an action based on the action instruction, the target attitude and position parameters are re-identified to dynamically correct the trajectory and realize closed-loop control.

Images