A method for autonomous operation of a robot in a space environment

Abstract

This invention belongs to the field of special robots and intelligent control, and discloses a method for autonomous robot operation in the space environment. This method addresses the challenge of high-precision and stable autonomous operation of robots in the complex lighting, communication-restricted, and microgravity environment of space. It achieves high-precision acquisition of the pose and motion state of non-cooperative targets by constructing a multi-sensor fusion stable perception mechanism adapted to the extreme space environment; it employs an edge computing-driven task autonomous planning strategy to achieve efficient autonomous decision-making and rapid response in space-based operations under communication-restricted conditions; and it proposes a force-position cooperative control method based on learning perturbation compensation to ensure high-precision and stable execution of space-based operations in a microgravity environment. The method proposed in this invention effectively improves the success rate of space robot operations and contributes to promoting the large-scale application of space manufacturing.

Description

Technical Field

[0001] This invention belongs to the field of robotics technology and relates to a method for autonomous robot operation in a space environment. Background Technology

[0002] With the rapid development of aerospace technology, human development and utilization of space resources is shifting from the traditional model of ground-based manufacturing and single-launch orbit insertion to a new stage of on-orbit manufacturing, on-orbit assembly, and on-orbit maintenance. Space manufacturing, as a key pillar of the aerospace system, can overcome the limitations of Earth's gravity and environment, enabling the preparation of high-purity materials, on-orbit assembly of large-scale structural components, and in-situ production and repair of spacecraft parts. This not only significantly reduces spacecraft launch costs but also builds sustainable and scalable space infrastructure. Its importance is increasingly prominent, making it one of the core technologies for future space development and deep space exploration.

[0003] However, space manufacturing faces complex conditions such as extreme temperature differences, strong radiation, and microgravity. Traditional manual on-orbit operation or remote ground control suffers from bottlenecks such as large delays, high risks, and low efficiency. As the core execution vehicle for space manufacturing, space robots possess autonomous perception, precise operation, and strong environmental adaptability. They can efficiently complete delicate tasks such as assembly, testing, and equipment maintenance in the manufacturing process under unattended or minimally manual operation conditions, providing indispensable technical support and safety guarantees for the implementation and large-scale application of space manufacturing.

[0004] The patent titled "A Visual Positioning Method for Satellite On-Orbit Refueling and Docking Services" (patent number CN 120374727 A) proposes an on-orbit refueling target positioning method based on multi-view image modeling and visual matching optimization. This method is used to achieve high-precision visual positioning of key structures such as docking holes. However, this method mainly focuses on pose solving at the static visual level and lacks the ability to continuously perceive the dynamic motion state of the satellite. The patent titled "A Method for Optimizing the Configuration of a Space Dual-Arm Robot Before Capture" (patent number CN 121245793 A) optimizes the robot's configuration by constructing a multi-objective optimization model before capture. This aims to reduce collision impacts during capture and ensure the operability of the closed-loop system after capture. However, this method primarily focuses on offline configuration planning and ideal dynamic performance evaluation before capture, without fully considering uncertainties and dynamic disturbances in the actual space environment. Therefore, this invention proposes a method for autonomous robot operation in the space environment. By constructing a multi-sensor stable perception mechanism adapted to the extreme space environment, continuous and high-precision perception of the pose and motion state of non-cooperative targets is achieved. Combined with an edge computing-driven task-level autonomous planning strategy and a force-position cooperative operation method based on learning perturbation compensation, the method ultimately enables safe, stable, and high-precision autonomous execution of typical on-orbit space operations such as refueling and component assembly. Summary of the Invention

[0005] The technical problems this invention aims to solve are as follows: traditional methods suffer from perception failures such as target loss and image corruption in harsh space environments with extreme temperature variations and complex lighting conditions; simultaneously, traditional teleoperation suffers from high latency and decision-making delays under limited space-to-ground communication conditions; furthermore, the weak constraints of traditional robot bases in microgravity environments lead to operational instability. To address these problems, this invention proposes an autonomous robot operation scheme for extreme space environments. By constructing a multi-sensor fusion stable perception mechanism, an edge computing-driven task-level autonomous planning strategy, and a force-position cooperative control method based on learning perturbation compensation, this scheme effectively achieves high-precision perception, autonomous decision-making, and stable operation execution of space robots against non-cooperative targets under communication-constrained and microgravity conditions.

[0006] The technical solution of the present invention: A method for autonomous robot operation in a space environment includes the following steps: The first step is to design and build an autonomous operating robot for the space environment; An autonomous robot designed for space environments is constructed, comprising a robotic arm, a vision camera, a force sensor, an inertial sensor, an edge computing chip, an active temperature control box, and an active light source. The active light source is fixed to the front end of the vision camera, the vision camera and the edge computing chip are fixed inside the active temperature control box, and the active temperature control box and the force sensor are fixed to the end of the robotic arm. The inertial sensor is fixed to the base of the robotic arm. The second step is to develop a multi-sensor stabilization compensation method for extreme space environments. To address the problem of visual camera measurement failure caused by extreme temperature changes and complex lighting in the space environment, a temperature compensation mechanism and an illumination imaging compensation mechanism that combine hardware control and algorithm correction are constructed. The temperature compensation mechanism is as follows: The active temperature control box consists of an insulated outer shell, an active heating component, a temperature sensor, and a temperature control circuit; the insulated outer shell is a double-layer insulated composite structure, with the outer layer made of polyurethane as the first layer of insulation barrier; an air cavity is reserved between the outer layer and the inner wall as the second layer of insulation structure; the active heating component uses the temperature sensor to provide feedback on the temperature inside the active temperature control box. When the temperature inside the active temperature control box exceeds the set temperature range, the temperature control circuit adjusts the power of the active heating component to stabilize the temperature within the set temperature range. Based on the temperature inside the active temperature control box fed back by the temperature sensor within a set temperature range, the measurement accuracy of the vision camera is corrected through a temperature compensation algorithm: (1) In the formula, Y comp The compensated visual camera measurements, Y raw These are the raw measurements from the vision camera. K 0 The zero-point temperature coefficient, K S The temperature coefficient of sensitivity. Y 00 The measured value is at the reference temperature. T The real-time temperature obtained from the temperature sensor. T 0 represents the normal operating reference temperature for the vision camera; The illumination imaging compensation mechanism is as follows: Images are acquired in real time using a visual camera, and the average brightness of the image is read. The calculation method is as follows: (2) in, M This represents the number of pixels in the height direction. N This refers to the number of pixels in the width direction. For the first in the image i Line number j The image brightness at each column position is adjusted; when the average image brightness is lower than the preset brightness, an active light source will be activated to increase the image brightness; when the average image brightness is higher than the preset brightness, an adaptive dynamic exposure adjustment algorithm will be used to adjust the exposure time. E t Reduce average image brightness The specific algorithm is as follows: (3) in, E t+1 This is the updated exposure time. I target For preset brightness, k This is the gain coefficient for exposure adjustment; The third step is high-precision perception of the pose and motion state of non-cooperative targets; When autonomous robots perform on-orbit operations in space environments, targets without visual markers for localization are defined as non-cooperative targets. A pose estimation method for non-cooperative targets is proposed: a parallax attention-enhanced keypoint detection network is constructed to detect the 2D keypoints on the feature image of the non-cooperative target captured by the visual camera corresponding to 3D keypoints on the non-cooperative target surface. This keypoint detection network is based on ResNet-18 and embeds a parallax attention module (DAM) during the feature extraction stage to enhance visual feature association. First, pooling is stopped when the feature image size decreases to H / 8×W / 8. The receptive field is kept constant by downsampling and dilated convolution, while the original fully connected layers are replaced with convolutional layers. Next, the acquired feature images are input into the disparity attention module, which generates a disparity attention weight map by calculating the disparity correlation between visual features. This weight map is then fused with the feature images to obtain a fused feature image, making the keypoint detection network more attentive to non-cooperative targets. Finally, after enhancing the disparity attention, the fused feature image undergoes skip connections, convolution, and upsampling until its size is restored to H×W. Finally, a 1×1 convolution outputs a unit vector and class probability, achieving 2D keypoint detection and matching. After completing the detection and matching of 2D keypoints, the 6D pose of the non-cooperative target is further solved: feature images of the non-cooperative target are acquired using a visual camera, and 2D keypoints are extracted from the feature images using a keypoint detection network; the extracted 2D keypoints are matched with 3D keypoints on the surface of the non-cooperative target to construct a 2D-3D keypoint correspondence; finally, the PnP algorithm is used to solve for the pose parameters of the constructed 2D-3D keypoint correspondence, and the 6D pose of the non-cooperative target is finally obtained. The PnP algorithm is as follows: (4) in, s For depth scale factor, ( u , v ) represents the coordinates of the 2D key points. X , Y , Z ( ) represents the 3D keypoint coordinates of the non-cooperative target. K For the camera intrinsic parameter matrix, [ R , t [ ] represents the 6D pose of the non-cooperative objective to be solved; The motion velocity of a non-cooperative target is estimated using a pose difference method: First, the translation and rotation parameters of the non-cooperative target in two adjacent frames are obtained, including the first... k Frame translation vector t k Rotation matrixR k With the k Translation vector of +1 frame t k+1 Rotation matrix R k+1 And calculate the time interval Δ between the two frames. t Subsequently passed t k+1 - R k+1 t k After eliminating the coordinate transformation error caused by the rotation of the non-cooperative target, the inter-frame translation displacement is obtained. This displacement is then divided by the time interval Δt to obtain the velocity of the non-cooperative target in the world coordinate system. v : (5) The fourth step is edge computing-driven task-level autonomous planning; The visual recognition network is integrated with a large language model and deployed in an edge computing chip to achieve physical information recognition of non-cooperative targets and autonomous task planning. The visual recognition network uses the ResNet-18 network built in the second step as the core network, takes the images acquired by the visual camera as input, and completes non-cooperative target detection, 6D pose estimation, and motion velocity estimation. It outputs the physical information of the non-cooperative target: target category, 6D pose, and motion velocity. The physical information of the non-cooperative target is transformed into a task state description in natural language form, represented in a unified text format of target category-6D pose-motion velocity, forming a task state description that can be directly understood by the large language model. The GPT-5 large language model is adopted as the core of autonomous task planning, and a task knowledge guide is introduced as a prior constraint. The task knowledge guide includes stage division prior knowledge, timing constraints, safety restrictions, and replanning rules. Among them, the stage division prior knowledge divides the on-orbit operation of the autonomous robot into four independent operation stages: target localization, proximity tracking, precision operation, and termination and withdrawal. The timing constraints specify the time window and execution interval of each operation stage. The safety restrictions cover the robot arm's movement speed threshold in the space environment, the relative distance safety margin between the robot arm and non-cooperative targets, and collision avoidance rules. The replanning rules include task pause and replanning in case of non-cooperative target's attitude change and abnormal robot arm movement. During the autonomous task planning process, the input to the GPT-5 large language model consists of four parts: the overall task objective. T goal Task Status Description S t Task Knowledge Guide G guide Historical execution information Ht , is represented as: (6) The GPT-5 large language model is based on the prior knowledge of stage division, temporal constraints, safety restrictions, and replanning rules provided by the task knowledge guide. Combined with contextual reasoning and temporal logic deduction capabilities, it comprehensively understands and decomposes the input information: First, it completes the understanding of the overall task goal and safety restrictions according to the task knowledge guide. Then, it decomposes the overall task goal into a sequence of sub-tasks that conform to the work process, are timely, and meet the requirements of safety and accuracy. At the same time, it generates the execution priority and stage switching conditions of each sub-task. The fifth step is to learn about force-potential stability operations driven by disturbance compensation in a microgravity environment. By constructing a Gaussian process regression model (GPR), real-time prediction of the recoil acceleration of the robotic arm base is achieved. First, the motion state information of the robotic arm is collected in real time: the motion parameters of the robotic arm are output by the motors and encoders at the joints of the robotic arm, including the joint angle, velocity, and acceleration; the force and torque at the end of the robotic arm are output by the force sensor fixed to the end of the robotic arm; the recoil acceleration of the robotic arm base is collected by the inertial sensor fixed to the base of the robotic arm; and the acceleration data of the robotic arm base under different motion states of the robotic arm are recorded to form a robotic arm motion state-robotic arm base acceleration dataset. Secondly, based on the robotic arm's motion state information x acceleration of the robotic arm base f ( x A Gaussian process regression model (GPR) is used to construct a base acceleration prediction model. The Gaussian process regression model is a probability-based regression model, and the base acceleration prediction model constructed based on this model is uniquely determined by the mean function and the covariance function. (7) In the formula, gp For Gaussian process regression model, m ( x ) is the mean function, k ( x , x ′) is the covariance function, used to measure the correlation between the motion state information of the robotic arm; Based on the constructed base acceleration prediction model, the real-time collected motion parameters of the robotic arm, the force and torque at the end of the robotic arm are input into the base acceleration prediction model, which predicts the base recoil acceleration of the robotic arm. Based on the predicted base recoil acceleration, the joint angle, speed and acceleration of the robotic arm are controlled in real time. By adjusting the motion state of the robotic arm, the base recoil acceleration of the robotic arm is effectively suppressed, ensuring that the robotic arm can complete the task stably and with high precision in a microgravity environment.

[0007] The beneficial effects of this invention are as follows: This invention addresses the challenge of autonomous, high-precision, and stable operation of robots in the space environment. It enables precise positioning and stable operation of non-cooperative targets without relying on ground control, providing key technical support for core tasks in space manufacturing processes such as component assembly, precision welding, and on-orbit maintenance. By improving the autonomy and mission success rate of autonomous robots in on-orbit operations, it helps promote the engineering implementation and large-scale application of space manufacturing technology, reduces the overall cost of space manufacturing, improves the utilization efficiency of space resources and assets, and promotes the high-quality development of the space manufacturing industry. It has significant engineering application value and substantial economic and social benefits. Attached Figure Description

[0008] Figure 1 A flowchart illustrating the on-orbit operation process of a space robot.

[0009] Figure 2 Flowchart for 6D pose of a non-cooperative target.

[0010] Figure 3 This is a structural diagram for ground-based experimental verification of a space-operating robot.

[0011] In the diagram: 1. Mobile platform, 2. Autonomous robot, 3. Force sensor, 4. Vision camera, 5. Non-cooperative target, 6. Suspension frame. Detailed Implementation

[0012] The specific embodiments of the present invention are described in detail below with reference to the technical solutions and accompanying drawings.

[0013] The first step is to design and build an autonomous operating robot for the space environment. The autonomous robot 2 designed for the space environment includes a robotic arm, a force sensor 3, a vision camera 4, an inertial sensor, an edge computing chip, an active temperature control box, and an active light source. The active light source is fixed to the front end of the vision camera 4. The vision camera 4 and the edge computing chip are fixed inside the active temperature control box, which, along with the force sensor 3, is fixed to the end effector of the robotic arm. The inertial sensor is fixed to the robotic arm base. The AUBO i5 robot, as the autonomous robot 2, has a repeatability accuracy of 0.02mm and is the core component for enabling on-orbit operations. The edge computing chip features high computing power and low power consumption, efficiently supporting the robot's physical information recognition and autonomous task planning during on-orbit operations. The force sensor 3 has an accuracy of 0.1%FS and a measurement range covering 500N and 15N. The N·m sensor provides force and torque information at the end effector of the robotic arm for motion status information; the vision camera 4 has a wide dynamic range and strong light interference resistance, with a resolution of up to 4K, and is used to achieve high-precision perception of the pose and motion status of non-cooperative targets; the inertial sensor has a measurement range of -16g to 16g and is used to provide the acceleration of the robotic arm base; the active temperature control box is used to ensure that the operating temperature of the vision camera 4 is within the set temperature range in the temperature compensation mechanism; the active light source has a power of 30W and is used to improve image brightness in the illumination imaging compensation mechanism.

[0014] The second step is to develop a multi-sensor stabilization compensation method for extreme space environments. To address the issue of decreased measurement accuracy or even failure of the visual camera 4 caused by extreme temperature variations and complex lighting conditions in the space environment, an environmental adaptive compensation method combining hardware control and algorithm correction is proposed. This method includes a temperature compensation mechanism and an illumination imaging compensation mechanism, which work together to improve the stability and reliability of the visual measurement system in complex environments.

[0015] In this embodiment, the temperature compensation mechanism consists of a heat-insulating shell, an active heating component, a temperature sensor, and a temperature control circuit. The heat-insulating shell is a double-layer heat-insulating composite structure, with the outer layer made of polyurethane as the first heat-insulating barrier. An air cavity is reserved between the outer layer and the inner wall as the second heat-insulating structure. The active heating component uses the temperature sensor to provide feedback on the temperature inside the active temperature control box. When the temperature inside the active temperature control box exceeds the set temperature range, the temperature control circuit adjusts the power of the active heating component to stabilize the temperature within the set temperature range, thereby ensuring that the vision camera 4 is in a stable working state.

[0016] Based on the temperature inside the active temperature control box fed back by the temperature sensor within a set temperature range, the measurement accuracy of the vision camera 4 is corrected through a temperature compensation algorithm: (8) In the formula, Y comp The compensated visual camera 4 measurements,Y raw These are the raw measurements from vision camera 4. K 0 The zero-point temperature coefficient, K S The temperature coefficient of sensitivity. Y 00 The measured value is at the reference temperature. T The real-time temperature obtained from the temperature sensor. T 0 represents the normal operating reference temperature for the vision camera 4; in practical applications, the zero-point temperature coefficient... K 0 With sensitivity temperature coefficient K S The zero-point temperature coefficient was determined by collecting measurement data from the vision camera 4 under different temperature conditions and then fitting the data. K 0 and sensitivity temperature coefficient K S This allows for effective compensation for temperature drift errors.

[0017] In this embodiment, the illumination imaging compensation mechanism is used to address the impact of strong light, shadows, and high contrast variations in the space environment on image quality. Specifically, the illumination imaging compensation mechanism is as follows: Images are acquired in real-time using the visual camera 4, and the average brightness of the images is read. The calculation method is as follows: (9) in, M This represents the number of pixels in the height direction. N This refers to the number of pixels in the width direction. I ( i , j ) is the first in the image i Line number j Image brightness at column positions; the average image brightness is calculated to reflect the current image exposure level.

[0018] Based on the relationship between the average image brightness and the preset target brightness, when the average image brightness is lower than the preset brightness, an active light source will be turned on to increase the image brightness; when the average image brightness is higher than the preset brightness, an adaptive dynamic exposure adjustment algorithm will be used to adjust the exposure time. E t Reduce average image brightness The specific algorithm is as follows: (10) in, E t+1 This is the updated exposure time. I target For preset brightness,k is the exposure adjustment gain coefficient. In actual operation, the exposure time is iteratively updated according to equation (10), and upper and lower limits of the exposure time are set to prevent overexposure or underexposure. When the error between the average brightness of the image and the preset target brightness is less than the preset threshold, the adjustment process is stopped, thereby achieving stable imaging.

[0019] The third step is high-precision perception of the motion posture of non-cooperative targets; like Figure 2 As shown, when the autonomous robot 2 performs on-orbit operations in the space environment, if the target does not have visual markers for localization, it is defined as a non-cooperative target 5. A pose estimation method for the non-cooperative target 5 is proposed: a parallax attention-enhanced keypoint detection network is constructed to detect the 3D keypoints on the surface of the non-cooperative target 5 and their corresponding 2D keypoints on the feature image of the non-cooperative target 5 acquired by the visual camera 4. This keypoint detection network is based on ResNet-18 and embeds a parallax attention module (DAM) during the feature extraction stage to enhance visual feature association. First, when the feature image size is reduced to H / 8×W / 8, Pooling downsampling is stopped, and dilated convolutions are used to maintain the receptive field. Simultaneously, the original fully connected layers are replaced with convolutional layers. Next, the acquired feature images are input into the disparity attention module. A disparity attention weight map is generated by calculating the disparity correlation between visual features, and this weight map is then fused with the feature images to obtain a fused feature image. This makes the keypoint detection network more focused on non-cooperative targets. Finally, after completing the disparity attention enhancement, skip connections, convolutions, and upsampling operations are performed on the fused feature image until the size is restored to H×W. Finally, a 1×1 convolution outputs a unit vector and class probability, achieving 2D keypoint detection and matching. After completing the detection and matching of 2D keypoints, the 6D pose of the non-cooperative target 5 is further solved: feature images of the non-cooperative target 5 are acquired through visual camera 4, and 2D keypoints are extracted from the feature images using a keypoint detection network; the extracted 2D keypoints are matched with 3D keypoints on the surface of the non-cooperative target 5 to construct a 2D-3D keypoint correspondence; finally, the PnP algorithm is used to solve the pose parameters of the constructed 2D-3D keypoint correspondence, and the 6D pose of the non-cooperative target 5 is finally obtained. The PnP algorithm is as follows: (11) in, s For depth scale factor, ( u , v ) represents the coordinates of the 2D key points. X , Y , Z ( ) represents the 3D keypoint coordinates of non-cooperative target 5. K For the camera intrinsic parameter matrix, [R , t [This refers to the 6D pose of the non-cooperative objective 5 to be solved;] To address the problem of estimating the motion velocity of a non-cooperative target 5, a pose difference method is employed. Specifically, the pose difference method is used to estimate the motion velocity of the non-cooperative target 5: firstly, the translation and rotation parameters of the non-cooperative target 5 in two adjacent frames are obtained, including the... k Frame translation vector t k Rotation matrix R k With the k Translation vector of +1 frame t k+1 Rotation matrix R k+1 And calculate the time interval Δ between the two frames. t Subsequently passed t k+1 - R k+1 t k After eliminating the coordinate transformation error caused by the rotation of the non-cooperative target 5, the inter-frame translation displacement is obtained. This displacement is then divided by the time interval Δt to obtain the velocity of the non-cooperative target 5 in the world coordinate system. v : (12) The fourth step is edge computing-driven task-level autonomous planning; In this embodiment, a visual recognition network and a large language model are integrated and deployed in an edge computing chip to achieve physical information recognition and autonomous task planning for a non-cooperative target 5. The visual recognition network uses the ResNet-18 network constructed in the second step as its core network, takes the image acquired by the visual camera 4 as input, and completes the detection of the non-cooperative target 5, 6D pose estimation, and motion velocity estimation, outputting the physical information of the non-cooperative target 5: target category, 6D pose, and motion velocity. The physical information of the non-cooperative target 5 is converted into a task state description in natural language form, represented in a unified text format of target category-6D pose-motion velocity, forming a task state description that can be directly understood by the large language model. In terms of task planning, this embodiment uses the GPT-5 large language model as the core of autonomous task planning and introduces a task knowledge guide as a prior constraint. The task knowledge guide includes stage division prior knowledge, timing constraints, safety restrictions, and replanning rules. Among them, the stage division prior knowledge divides the on-orbit operation of the autonomous robot 2 into four independent operation stages: target localization, proximity tracking, precision operation, and termination and withdrawal. The timing constraints specify the time window and execution interval for each operation stage. The safety restrictions cover the robot arm's movement speed threshold in the space environment, the relative distance safety margin between the robot arm and the non-cooperative target 5, and collision avoidance rules. The replanning rules include the sudden change in the attitude of the non-cooperative target 5 and the task suspension and replanning in the case of abnormal robot arm movement. During the autonomous task planning process, the input to the GPT-5 large language model consists of four parts: the overall task objective. T goal Task Status Description S t Task Knowledge Guide G guide Historical execution information H t , is represented as: (13) The GPT-5 large language model is based on the prior knowledge of stage division, temporal constraints, safety restrictions, and replanning rules provided by the task knowledge guide. Combined with contextual reasoning and temporal logic deduction capabilities, it comprehensively understands and decomposes the input information: First, it completes the understanding of the overall task goal and safety restrictions according to the task knowledge guide. Then, it decomposes the overall task goal into a sequence of sub-tasks that conform to the work process, are timely, and meet the requirements of safety and accuracy. At the same time, it generates the execution priority and stage switching conditions of each sub-task.

[0020] The fifth step is to learn about force-potential stability operations driven by disturbance compensation in a microgravity environment. In this embodiment, a Gaussian process regression model (GPR) is constructed to achieve real-time prediction of the recoil acceleration of the robotic arm base. First, the motion state information of the robotic arm is collected in real time: the motion parameters of the robotic arm are output by the motor and encoder at the joints of the robotic arm, including the joint angle, velocity, and acceleration of the robotic arm; the force and torque at the end of the robotic arm are output by the force sensor 3 fixed to the end of the robotic arm; the recoil acceleration of the robotic arm base is collected by the inertial sensor fixed to the base of the robotic arm; and the acceleration data of the robotic arm base under different motion states of the robotic arm are recorded to form a robotic arm motion state-robotic arm base acceleration dataset. Secondly, based on the robotic arm's motion state information x acceleration of the robotic arm base f ( xA Gaussian process regression model (GPR) is used to construct a base acceleration prediction model. The Gaussian process regression model is a probability-based regression model, and the base acceleration prediction model constructed based on this model is uniquely determined by the mean function and the covariance function. (14) In the formula, gp It is a Gaussian process regression model; m ( x () is the mean function; k ( x , x ′) is the covariance function, used to measure the correlation between the motion state information of the robotic arm. It can effectively capture the smooth change characteristics between the motion state of the robotic arm and the acceleration of the base. Its expression is as follows: (15) in, For signal variance, l For length scale, This represents the Euclidean distance between the inputs of the robotic arm's motion state information.

[0021] After the base acceleration prediction model is built, the real-time collected data on the robot arm joint angles, velocities, accelerations, end effector forces, and torques are input into the model to predict the recoil acceleration of the robot arm base. Subsequently, the robot arm joint angles, velocities, and accelerations are adjusted in real time based on the predicted base recoil acceleration, thereby effectively suppressing the robot arm base recoil, accelerating the stability of the robot arm in a microgravity environment, and ensuring that it can complete on-orbit tasks with high precision.

[0022] Step 6: Ground-based experimental verification. like Figure 3 As shown, to verify the effectiveness and stability of the autonomous robot operation method for the space environment of the present invention in actual operation scenarios, a ground verification experimental system was further constructed. A composite robot was used as a ground-based equivalent of the space operation robot. The composite robot includes a mobile platform 1 and an autonomous operation robot 2 to simulate the on-orbit operation state of a freely floating robot in space. The non-cooperative target 5 is suspended by a suspension frame 6 to counteract the effects of gravity, placing it in a quasi-free state within the experimental space, approximately simulating the motion characteristics of the target satellite operating in the microgravity environment of space.

[0023] like Figure 1As shown, during the experiment, the non-cooperative target 5 was first placed in different initial positions and postures in the experimental space. Then, the autonomous robot 2 started from different initial positions, and the stable operation of the vision camera 4 was ensured through temperature compensation mechanism and illumination imaging compensation mechanism. Second, the vision camera 4 was used to collect feature images of the non-cooperative target 5, and the key points in the feature images were detected by the key point detection network. The 6D pose of the non-cooperative target 5 could be obtained by the PNP algorithm. Furthermore, the pose difference method was used to realize the real-time estimation of the motion velocity of the non-cooperative target 5. Third, the large language model was used to realize the task-level autonomous planning of the autonomous robot 2's on-orbit operation. Finally, the Gaussian process regression model was used to construct the base acceleration prediction model. The robot's joint motion can be controlled by collecting the robot's motion state information to realize the robot's stable on-orbit operation.

[0024] Through repeated experiments, the success rate of the non-cooperative target 5 under different initial positions and attitudes, and the autonomous robot 2 starting from different positions, was statistically analyzed to evaluate the feasibility of the proposed method. Experimental results show that the non-cooperative target 5 identification accuracy reaches 100%, and the success rate is 85%, fully verifying the feasibility of the proposed autonomous robot operation method in space.

Claims

1. A method for autonomous robot operation in a space environment, characterized in that, Includes the following steps: The first step is to design and build an autonomous operating robot for the space environment; The second step is to develop a multi-sensor stabilization compensation method for extreme space environments. To address the problem of visual camera measurement failure caused by extreme temperature changes and complex lighting in the space environment, a temperature compensation mechanism and an illumination imaging compensation mechanism that combine hardware control and algorithm correction are constructed. The third step is high-precision perception of the pose and motion state of non-cooperative targets; The fourth step is edge computing-driven task-level autonomous planning; The fifth step is to learn about force-position stabilization operations in microgravity environments driven by disturbance compensation.

2. The method for autonomous robot operation in a space environment according to claim 1, characterized in that, The specific implementation process of the first step is as follows: An autonomous robot designed for space environments is constructed, comprising a robotic arm, a vision camera, a force sensor, an inertial sensor, an edge computing chip, an active temperature control box, and an active light source. The active light source is fixed to the front of the vision camera, the vision camera and the edge computing chip are fixed inside the active temperature control box, and the active temperature control box and the force sensor are fixedly connected to the end of the robotic arm. The inertial sensor is fixed to the base of the robotic arm.

3. The method for autonomous robot operation in a space environment according to claim 2, characterized in that, The specific implementation process of the second step is as follows: The temperature compensation mechanism is as follows: The active temperature control box consists of an insulated outer shell, an active heating component, a temperature sensor, and a temperature control circuit; the insulated outer shell is a double-layer insulated composite structure, with the outer layer made of polyurethane as the first layer of insulation barrier; an air cavity is reserved between the outer layer and the inner wall as the second layer of insulation structure; the active heating component uses the temperature sensor to provide feedback on the temperature inside the active temperature control box. When the temperature inside the active temperature control box exceeds the set temperature range, the temperature control circuit adjusts the power of the active heating component to stabilize the temperature within the set temperature range. Based on the temperature inside the active temperature control box fed back by the temperature sensor within a set temperature range, the measurement accuracy of the vision camera is corrected through a temperature compensation algorithm: In the formula, Y comp The compensated visual camera measurements, Y raw These are the raw measurements from the vision camera. K 0 The zero-point temperature coefficient, K S The temperature coefficient of sensitivity. Y 00 The measured value is at the reference temperature. T The real-time temperature obtained by the temperature sensor. T 0 represents the normal operating reference temperature for the vision camera; The illumination imaging compensation mechanism is as follows: Images are acquired in real time using a visual camera, and the average brightness of the image is read. The calculation method is as follows: in, M This represents the number of pixels in the height direction. N This refers to the number of pixels in the width direction. For the first in the image i Line number j The image brightness at each column position is adjusted; when the average image brightness is lower than the preset brightness, an active light source will be activated to increase the image brightness; when the average image brightness is higher than the preset brightness, an adaptive dynamic exposure adjustment algorithm will be used to adjust the exposure time. E t Reduce average image brightness The specific algorithm is as follows: in, E t+1 This is the updated exposure time. I target For preset brightness, k This is the exposure adjustment gain coefficient.

4. The method for autonomous robot operation in a space environment according to claim 3, characterized in that, The specific implementation process of the third step is as follows: When autonomous robots perform on-orbit operations in space environments, targets without visual markers for localization are defined as non-cooperative targets. A pose estimation method for non-cooperative targets is proposed: a parallax attention-enhanced keypoint detection network is constructed to detect the 2D keypoints on the feature image of the non-cooperative target captured by the visual camera corresponding to 3D keypoints on the non-cooperative target surface. This keypoint detection network is based on ResNet-18 and embeds a parallax attention module (DAM) during the feature extraction stage to enhance visual feature association. First, pooling is stopped when the feature image size decreases to H / 8×W / 8. The receptive field is kept constant by downsampling and dilated convolution, while the original fully connected layers are replaced with convolutional layers. Next, the acquired feature images are input into the disparity attention module, which generates a disparity attention weight map by calculating the disparity correlation between visual features. This weight map is then fused with the feature images to obtain a fused feature image, making the keypoint detection network more attentive to non-cooperative targets. Finally, after enhancing the disparity attention, the fused feature image undergoes skip connections, convolution, and upsampling until its size is restored to H×W. Finally, a 1×1 convolution outputs a unit vector and class probability, achieving 2D keypoint detection and matching. After completing the detection and matching of 2D keypoints, the 6D pose of the non-cooperative target is further solved: feature images of the non-cooperative target are acquired using a visual camera, and 2D keypoints are extracted from the feature images using a keypoint detection network; the extracted 2D keypoints are matched with 3D keypoints on the surface of the non-cooperative target to construct a 2D-3D keypoint correspondence; finally, the PnP algorithm is used to solve for the pose parameters of the constructed 2D-3D keypoint correspondence, and the 6D pose of the non-cooperative target is finally obtained. The PnP algorithm is as follows: in, s For depth scale factor, ( u , v ) represents the coordinates of the 2D key points. X , Y , Z ( ) represents the 3D keypoint coordinates of the non-cooperative target. K For the camera intrinsic parameter matrix, [ R , t [ ] represents the 6D pose of the non-cooperative objective to be solved; The motion velocity of a non-cooperative target is estimated using a pose difference method: First, the translation and rotation parameters of the non-cooperative target in two adjacent frames are obtained, including the first... k Frame translation vector t k Rotation matrix R k With the k Translation vector of +1 frame t k+1 Rotation matrix R k+1 And calculate the time interval Δ between the two frames. t Subsequently passed t k+1 - R k+1 t k After eliminating the coordinate transformation error caused by the rotation of the non-cooperative target, the inter-frame translation displacement is obtained. This displacement is then divided by the time interval Δt to obtain the velocity of the non-cooperative target in the world coordinate system. v : 。 5. The method for autonomous robot operation in a space environment according to claim 4, characterized in that, The specific implementation process of the fourth step is as follows: The visual recognition network is integrated with a large language model and deployed in an edge computing chip to achieve physical information recognition of non-cooperative targets and autonomous task planning. The visual recognition network uses the ResNet-18 network built in the second step as the core network, takes the images acquired by the visual camera as input, and completes non-cooperative target detection, 6D pose estimation, and motion velocity estimation. It outputs the physical information of the non-cooperative target: target category, 6D pose, and motion velocity. The physical information of the non-cooperative target is transformed into a task state description in natural language form, represented in a unified text format of target category-6D pose-motion velocity, forming a task state description that can be directly understood by the large language model. The GPT-5 large language model is adopted as the core of autonomous task planning, and a task knowledge guide is introduced as a prior constraint. The task knowledge guide includes stage division prior knowledge, timing constraints, safety restrictions, and replanning rules. Among them, the stage division prior knowledge divides the on-orbit operation of the autonomous robot into four independent operation stages: target localization, proximity tracking, precision operation, and termination and withdrawal. The timing constraints specify the time window and execution interval of each operation stage. The safety restrictions cover the robot arm's movement speed threshold in the space environment, the relative distance safety margin between the robot arm and non-cooperative targets, and collision avoidance rules. The replanning rules include task pause and replanning in case of non-cooperative target's attitude change and abnormal robot arm movement. During the autonomous task planning process, the input to the GPT-5 large language model consists of four parts: the overall task objective. T goal Task Status Description S t Task Knowledge Guide G guide Historical execution information H t , is represented as: The GPT-5 large language model is based on the prior knowledge of stage division, temporal constraints, safety restrictions, and replanning rules provided by the task knowledge guide. Combined with contextual reasoning and temporal logic deduction capabilities, it comprehensively understands and decomposes the input information: First, it completes the understanding of the overall task goal and safety restrictions according to the task knowledge guide. Then, it decomposes the overall task goal into a sequence of sub-tasks that conform to the work process, are timely, and meet the requirements of safety and accuracy. At the same time, it generates the execution priority and stage switching conditions of each sub-task.

6. The method for autonomous robot operation in a space environment according to claim 5, characterized in that, The specific implementation process of step five is as follows: By constructing a Gaussian process regression model (GPR), real-time prediction of the recoil acceleration of the robotic arm base is achieved. First, the motion state information of the robotic arm is collected in real time: the motion parameters of the robotic arm are output by the motors and encoders at the joints of the robotic arm, including the joint angle, velocity, and acceleration; the force and torque at the end of the robotic arm are output by the force sensor fixed to the end of the robotic arm; the recoil acceleration of the robotic arm base is collected by the inertial sensor fixed to the base of the robotic arm; and the acceleration data of the robotic arm base under different motion states of the robotic arm are recorded to form a robotic arm motion state-robotic arm base acceleration dataset. Secondly, based on the robotic arm's motion state information x With the acceleration of the robotic arm base f ( x A Gaussian process regression model (GPR) is used to construct a base acceleration prediction model. The Gaussian process regression model is a probability-based regression model, and the base acceleration prediction model constructed based on this model is uniquely determined by the mean function and the covariance function. In the formula, gp For Gaussian process regression model, m ( x ) is the mean function, It is a covariance function used to measure the correlation between the motion state information of the robotic arm; Based on the constructed base acceleration prediction model, the real-time collected motion parameters of the robotic arm, the force and torque at the end of the robotic arm are input into the base acceleration prediction model, which predicts the base recoil acceleration of the robotic arm. Based on the predicted base recoil acceleration, the joint angle, speed and acceleration of the robotic arm are controlled in real time. By adjusting the motion state of the robotic arm, the base recoil acceleration of the robotic arm is effectively suppressed, ensuring that the robotic arm can complete the task stably and with high precision in a microgravity environment.

Images