A method and system for dexterous grasping control of mobile robots based on tactile feedback

Abstract

The present application relates to a kind of based on haptic feedback mobile robot dexterous grasping control method and system, it is related to the field of mobile robot, including: the point cloud data of target object is acquired, multiple candidate grasping poses are generated using conditional variational autoencoder, candidate grasping pose includes end effector position, pose and dexterous hand joint configuration;Optimal grasping guide is selected from candidate grasping pose based on physical encumbrance index, physical encumbrance index includes grasping width coverage and grasping depth coverage;Optimal grasping guide, robot body state, object relative pose and tactile contact signal are input into reinforcement learning strategy based on proximal policy optimization, generate whole body collaborative control instruction to drive robot to execute grasping, and real-time adjustment grip gesture according to tactile signal during grasping process.The present application realizes the whole body collaborative robust grasping of chassis, manipulator and dexterous hand of mobile robot in high-speed dynamic scene.

Description

Technical Field

[0001] This invention relates to the field of mobile robots, and in particular to a method and system for dexterous grasping control of mobile robots based on haptic feedback. Background Technology

[0002] Achieving dexterous grasping in mobile robots is an important technology in the field of robot control. The realization of high-speed response, full-body coordination and scene generalization during movement requires high-quality dynamic perception and efficient control and decision-making methods.

[0003] However, existing research still has significant limitations: 1. Current dexterous grasping research focuses on dexterous grasping with stationary robots that do not require movement, thus limiting the operational range of dexterous grasping. 2. Existing research on mobile grasping is mostly focused on parallel claw robots. The simple structure of parallel claw robots cannot perform fine manipulation, and the methods are difficult to transfer to multi-fingered dexterous hands, limiting their application in grasping complex objects. 3. Existing grasping research mainly uses tactile sensing for static grasping, making it difficult to achieve real-time grip adjustment and anti-slip during high-speed movement.

[0004] Therefore, there is an urgent need for a method that can integrate tactile information to achieve full-body control of movement and grasping using dexterous hands. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for dexterous grasping control of mobile robots based on haptic feedback, so as to solve the problems of coordinated control and grasping stability in mobile grasping tasks.

[0006] To achieve the above objectives, the present invention provides a method for dexterous grasping control of a mobile robot based on haptic feedback, comprising the following steps:

[0007] The point cloud data of the target object is acquired, and multiple candidate grasping poses are generated using a conditional variational autoencoder. The candidate grasping poses include the position and orientation of the end effector and the dexterous hand joint configuration. The optimal grasping guide is selected from candidate grasping poses based on the physical bounding degree index, which includes grasping width coverage and grasping depth coverage. The optimal grasping guidance, robot body state, object relative pose, and tactile contact signals are input into a reinforcement learning strategy based on proximal strategy optimization to generate whole-body collaborative control commands to drive the robot to perform grasping, and the grasping gestures are adjusted in real time according to tactile signals during the grasping process.

[0008] Preferably, the steps include: Geometric features are extracted from the normalized object point cloud using a PointNet encoder. Latent variables are sampled from a Gaussian distribution and concatenated with geometric features before being input into the decoder to reconstruct the grasping pose; By sampling different latent variables multiple times, several candidate grasping postures with differences in spatial distribution and hand configuration are generated.

[0009] Preferably, this includes preliminary filtering and physical enclosure assessment: The initial filtering includes executability filtering and collision detection filtering. The executability filtering excludes non-forward-facing grasping based on the forward grasping space, while the collision detection filtering excludes candidate solutions where the end effector has descended below the lowest point of the object. The physical enclosure assessment calculates a comprehensive quality score based on the grasp width coverage and grasp depth coverage, and selects the grasp pose with the highest quality score as the optimal grasp guide.

[0010] Preferably, the calculation of the crawl width coverage includes: Define the vector connecting the thumb point and the finger point as the finger grip axis, and calculate the projection interval of the object's point cloud along this axis. The ratio of the intersection length of the finger's grasping area and the object's projection area to the total length of the object's projection area is calculated and used as the grasping width coverage of that finger.

[0011] Preferably, the calculation of the capture depth coverage includes: Define the projection from the center of the palm to the finger position in the direction perpendicular to the palm as the grip depth interval, and calculate the projection interval of the object point cloud in this depth direction. The ratio of the intersection length of the grip depth range of the finger and the palm with the object projection range to the total length of the object projection range is used as the grip depth coverage of that finger.

[0012] The preferred formula for calculating the overall quality score is:

[0013] in, Let the grasping width coverage of the k-th finger be . Let k be the grasping depth coverage of the k-th finger. This represents the gripping depth coverage of the thumb.

[0014] Preferably, the input observation vectors for the reinforcement learning strategy include: The linear and angular velocities of the chassis, the angles of each joint of the robotic arm, and the angles of each joint of the dexterous hand constitute the body state; The relative pose information is formed by the three-dimensional coordinates of the object in the robot coordinate system and its relative distance to the dexterous hand; Binary contact signals are transmitted by tactile sensors installed in the fingertips, finger pads, and palm areas; And optimal crawling guidance information.

[0015] Preferably, adjusting the grip gesture in real time based on tactile signals includes: When the tactile sensor on a finger does not detect a touch, the finger joint where the sensor is installed automatically tightens to form a tighter contact; When an object is detected to be sliding, the finger joint torque is dynamically adjusted to maintain a stable grip.

[0016] Preferably, this includes setting a phased reward function: At a distance of 0.5 meters from the object, the robot is encouraged to approach the target at high speed and maintain the height of the dexterous hand by rewarding the arm's operating radius, chassis movement speed, chassis movement direction, and height. Within 0.1 meters of the object, the robot is encouraged to precisely align and close its dexterous hand to form multi-point contact through grasping posture alignment rewards, dexterous hand movement rewards, stable grasping rewards, and tactile rewards.

[0017] The technical solution of the present invention also provides a mobile robot dexterous grasping control system based on tactile feedback, which can execute the above-described mobile robot dexterous grasping control method based on tactile feedback.

[0018] Compared with the prior art, the present invention has the following beneficial technical effects: 1. Significantly improves the success rate and stability of mobile robot grasping in high-speed dynamic scenarios. Based on a selection mechanism of grasping width and depth coverage, the most stable grasping guidance can be selected quickly and robustly from a geometric perspective, effectively resisting the impact and object slippage caused by high-speed contact. Combined with real-time adjustment of gripping force using binary tactile signals, the robot can adjust its grasping posture in real time to achieve stable grasping.

[0019] 2. Deep coordination between full-body movement and dexterous manipulation in high-speed motion was achieved, solving the challenge of joint control of mobile robots and multi-finger dexterous hands. A unified reinforcement learning strategy directly outputs full-body coordinated control commands, enabling the robot to accurately execute the entire process of approach, alignment, and grasping while maintaining high-speed movement, overcoming the problem of disconnect between movement and manipulation in traditional methods.

[0020] 3. Enhances the system's generalization ability and deployment convenience for unknown objects and environments. The generated model only requires object point clouds to produce diverse grasping capabilities, and the selection criteria do not depend on the complete geometric model, enabling the system to directly handle unseen objects and providing an end-to-end solution for complex dynamic operations of mobile robots. Attached Figure Description

[0021] Figure 1This is an overall overview of the dexterous grasping control method and system for mobile robots based on haptic feedback according to the present invention; Figure 2 This is a schematic diagram of the grasping generation process in a mobile robot dexterous grasping control method and system based on tactile feedback, according to the present invention. Figure 3 This is a schematic diagram of a tactile feedback-based dexterous grasping control method and system for mobile robots, and a grasping and filtering process according to the present invention. Detailed Implementation

[0022] 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.

[0023] This invention discloses a method and system for dexterous grasping control of a mobile robot based on tactile feedback. Its overall architecture covers three closely connected stages: grasp generation, grasp selection, and grasp implementation. By deeply integrating tactile perception and reinforcement learning technologies, it enables the mobile robot to perform full-body coordinated grasping operations in high-speed dynamic scenarios.

[0024] Specifically, firstly, a variety of candidate grasping poses are generated based on the object point cloud using a conditional variational autoencoder. Then, the most geometrically stable grasping guidance is selected using a quantitative index based on physical enclosedness. Finally, a reinforcement learning strategy that integrates tactile feedback is used to coordinate the full-body movement of the robot chassis, robotic arm, and multi-fingered dexterous hand. During the approach to the target object, the grasping posture is adjusted in real time to counteract the impact and slippage caused by high-speed contact, thereby completing a stable and reliable grasping task.

[0025] In the grasping generation stage, the system adopts a conditional variational autoencoder as the core generation model. This model takes the object point cloud as the conditional input and controls the generation diversity by sampling latent variables from a Gaussian distribution. Finally, the decoded output includes the complete grasping pose, including the position and attitude of the end effector and the dexterous hand joint configuration.

[0026] To ensure the model's generalization ability, a large-scale synthetic dataset containing 478,200 validated grasping poses was first constructed. This dataset covers 4,782 object instances, with each instance containing 100 different grasping poses. During model training, the encoder infers the posterior distribution of latent variables based on object point cloud features and successful grasping poses, while the decoder reconstructs grasping poses using latent codes sampled from the prior distribution and point cloud features. The latent space is normalized by maximizing the lower bound of evidence while constraining the KL divergence, enabling the model to learn to generate diverse and stable grasping candidates.

[0027] Furthermore, the training process uses PointNet to extract geometric features from the normalized point cloud of the target object as conditional input, and combines these with latent variables sampled from a Gaussian distribution to train the network to maximize the following lower bound of evidence: (1) Among them, encoder Based on object point cloud features and crawling Inferring latent variables The posterior distribution, decoder Using from prior distribution Potential code for mid-sampling and point cloud features Generate grasping poses. By normalizing the latent space by minimizing the KL divergence during training, while maximizing the probability of generating grasps, the model learns to generate diverse and stable grasping candidates from the input.

[0028] During the inference phase, the decoder receives the concatenated representation of point cloud features and latent variables randomly sampled from a Gaussian distribution as input, reconstructs the grasping pose from it, and generates 150 candidate grasping poses located at different spatial locations for each target object to ensure sufficient grasping diversity to cover a variety of possible grasping configurations.

[0029] After entering the grasping and filtering stage, the system quickly selects the optimal guidance most suitable for high-speed dynamic execution based on a large number of generated candidate grasps. This stage first performs preliminary filtering based on kinematics and collision to eliminate infeasible or inefficient grasps. Then, it proposes a grasping quality quantification index based on physical enclosedness to select the geometrically most stable grasping pose from the remaining candidates.

[0030] The initial filtering includes two levels: executability filtering and collision detection filtering. Executability filtering excludes grasping actions that require the robotic arm to bypass objects or move along long trajectories by defining a forward grasping space. Specifically, based on the relative posture between the robot and the target object, a grasping cone is established with the target's normal vector pointing to the robot as the reference. Forward grasping within this space has the shortest motion path and minimal joint motion, while non-forward grasping actions with the end effector rotating more than 90 degrees are filtered out by the system due to significantly increased execution time. Collision detection filtering excludes candidate solutions where the end effector would descend below the lowest point of the object to prevent collisions with the support surface.

[0031] After initial filtering, the system calculates the grasp width coverage (GWC) and grasp depth coverage (GDC) for each candidate grasp. The grasp width coverage quantifies the degree to which the grasping interval vector formed between the fingers and thumb covers the maximum effective width of the object along this vector. Based on the mechanical principle that grasping stability is improved when the contact points are located on opposite sides of the object, the system defines the finger grasping axis as the point connecting the thumb. and finger points vector ,make Indicate the corresponding unit direction, and let and These represent the scalar projections of the two fingertips onto this axis. The maximum effective width of an object is defined as the object's point cloud on... The distance between the two farthest points when projected in a direction is represented in this one-dimensional projection space, and the point cloud is labeled as... Then the fingers The GWC is: (2) Among them, the form is " " represents the length of a one-dimensional interval, "" indicates the intersection in the scalar space.

[0032] Grasp depth coverage quantifies the extent to which the gripping area formed between the fingers and the palmar reference plane covers the maximum effective depth of the object. A larger depth value indicates that the object is positioned deeper within the hand's natural gripping range. This deeper gripping position significantly increases the likelihood of contact with the fingertips and a wider area of ​​the palm (such as the fingertips and palm). Therefore, it facilitates multi-point distributed contact, thereby enhancing grip robustness.

[0033] vector Defined as the projection of the vector from the center of the palm to the finger position onto a direction perpendicular to the palm surface, representing the depth that the finger can reach, let... The unit normal vector of the palm plane is used to calculate the scalar depth along this direction. and In formula (3), the symbol Refers to fingers The one-dimensional depth range of the hand on the vertical palmar surface represents the maximum theoretical grip depth that the finger structure can provide under a specific grip posture. Defined as an object point cloud When projecting in a direction, the interval formed by the distance between the two farthest points. Then, the GDC of finger k is: (3) Overall grip quality considers coverage information in both width and depth dimensions, with the grip depth of the thumb being a key factor affecting overall grip stability. The grip method with the highest quality score is selected as a guideline. (4) During the grasping phase, the system integrates the robot's body state, the object's relative pose, tactile contact signals, and the filtered grasping guidance information. It then coordinates the robot's whole-body movement to complete the grasping task through a reinforcement learning strategy based on the proximal policy optimization (PPO) algorithm.

[0034] Specifically, the input to the strategy model includes the body state consisting of the linear and angular velocities of the chassis, the joint angles of the robotic arm and the joint angles of the dexterous hand, the relative pose information consisting of the coordinates of the object in the robot coordinate system and the distance to the dexterous hand, the contact signals transmitted by the tactile sensing devices installed in the fingertips, finger pads and palm areas of each finger, and the grasping guidance information including the position of the end effector, the rotational posture of the hand and the joint configuration of the dexterous hand.

[0035] The model output includes linear and angular velocity control commands for the chassis, joint angle control commands for the robotic arm, and joint angle control commands for the dexterous hand, achieving unified and coordinated control of the entire robot, including the moving chassis, robotic arm, and multi-fingered dexterous hand. During the grasping process, the system introduces a tactile feedback mechanism to adjust the gripping force in real time. When the tactile sensor on a certain finger does not detect touch, the system automatically tightens the finger joint where the sensor is installed to form a tighter contact.

[0036] To guide the strategy in learning the complete grasping process, the system is designed with a phased reward function: At a distance of 0.5 meters from the object, the arm's operating radius reward encourages significant arm extension to improve responsiveness to sudden changes in movement. Chassis movement speed and direction rewards promote high-speed approach to the target, where the chassis movement direction is the target's direction of motion, i.e., the point the robot initially points towards along the initial horizontal heading plus the offset of the robotic arm's operating radius. Simultaneously, height rewards encourage the dexterous hand to maintain a high distance from the ground. Within 0.1 meters of the object, the system uses grasping posture alignment rewards to encourage the robot to align its grasping posture with the grasping guide. Dexterous hand movement rewards promote rapid hand movement to complete the grasping closure. Stable grasping rewards extend the time the object remains grasped without falling. Finally, tactile rewards encourage the triggering of more tactile sensors to form multi-point contact.

[0037] This reinforcement learning strategy, guided by the aforementioned reward signals, enables the robot to accurately execute the entire process of approaching, aligning, and grasping while maintaining high-speed movement, overcoming the problem of the disconnect between motion and operation in traditional methods.

[0038] The three stages collaborate closely through data and control flows to form a complete grasping solution. The 150 candidate grasping poses generated in the grasp generation stage are passed as input to the grasping screening stage. After executability filtering, collision detection filtering, and physical enclosing-based quality assessment, the grasping pose with the highest quality score is guided to the grasping implementation stage. In the grasping implementation stage, a reinforcement learning strategy uses the grasping guide as a target pose reference, combining it with real-time perceived state information to generate control commands. Simultaneously, the haptic feedback module continuously monitors the contact state during execution and triggers real-time adjustments to the finger joints.

[0039] This collaborative mechanism enables the system to start from the input of an object point cloud, process it through three levels of generation, filtering, and execution, and finally output coordinated full-body control commands, achieving end-to-end mapping from perception to action. In simulation platform deployment, the system acquires the robot's body state, the object's relative pose, and tactile contact signals through the simulation interface. Based on the object's complete point cloud, it generates candidate grasping objects and filters out guides, controlling the robot's full-body movements, including its dexterous hand. During grasping, the system adjusts the grasping gesture based on tactile information. In real-world deployment, the system acquires the robot's body state through the ROS system, captures object point clouds using an RGBD camera, generates candidate grasping objects and filters out guides based on the point cloud, controls the robot to approach the object through the ROS system, and uses its dexterous hand to complete the grasping operation. Similarly, during grasping, the system uses feedback signals from tactile sensors to adjust the grasping gesture to achieve stable grasping, thus verifying the effectiveness and generalization ability of the method in simulation and real-world environments.

[0040] Reference Figure 1This figure illustrates the complete execution flow and data flow of the haptic feedback-based dexterous grasping control method for mobile robots provided by this invention. Specifically, the system first determines the optimal grasping guidance g based on the object point cloud through the aforementioned grasping generation and filtering mechanism. This grasping guidance information, the object's three-dimensional coordinates Pos in the robot coordinate system, the relative distance Dist between the object and the dexterous hand, and the real-time angles q of each joint of the robotic arm are then used to determine the optimal grasping guidance g. t Real-time angles c of each joint of the dexterous hand t And the current linear velocity and angular velocity v of the moving chassis. t The observation vectors are combined to form the input to the policy model. Based on these observation vectors, the policy model infers and outputs full-body control motion commands for the chassis, robotic arm, and dexterous hand. These motion commands are transmitted to the simulator or physical robot actuators to drive the system's movement. During the model training phase, the system further calculates the aforementioned phased reward function based on the current state and execution results, and updates the policy network parameters through a proximal policy optimization algorithm to optimize subsequent decision-making performance.

[0041] Reference Figure 2 This figure illustrates a specific implementation scheme for a grasping proposal generation mechanism based on a conditional variational autoencoder. The system first uses a PointNet encoder to extract geometric features from the input object point cloud, obtaining point cloud features F with permutation invariance. Simultaneously, it randomly samples feature vector z from a standard Gaussian distribution p(z) to obtain a feature vector z in the latent space. The point cloud features F and the latent vector z are concatenated and input into the decoder network, where the grasping pose is reconstructed through a conditional generation process. By sampling different latent vectors z multiple times, the system can generate up to 150 candidate grasping poses for a single target object, each with significant differences in spatial distribution and hand configuration, thus forming a diverse set of grasping candidates.

[0042] Reference Figure 3 The diagram illustrates the specific processing flow and evaluation mechanism of the grasping and filtering phase. The system sequentially filters the aforementioned candidate grasping sets through two preliminary layers: executability filtering and collision detection filtering, eliminating grasping schemes that are infeasible or have potential collisions with the support surface. For the remaining candidates that pass the preliminary filtering, the system calculates their grasping width coverage (GWC) and grasping depth coverage (GDC) indices, and calculates the overall quality score for each candidate. Finally, the system selects the grasping posture with the highest quality value as the optimal grasping guide, which is used to guide the subsequent reinforcement learning control strategy to execute the grasping operation.

[0043] The method of this invention is deployed on a robot built on a simulation platform. The simulation platform acquires the robot's body state, the relative pose of the object, and tactile contact signals. A large number of candidate grasping objects are generated from the complete point cloud of the object, and grasping guidance is selected. The robot's whole-body movement, including the dexterous hand, is controlled to grasp objects of various shapes. During the grasping process, the grasping gesture is adjusted through tactile information to achieve stable grasping.

[0044] The method of this invention is deployed on a robot platform built in the real world. The robot body state is obtained through the ROS system, and the object point cloud is obtained by taking pictures through the RGBD camera. A large number of candidate grasping objects are generated based on the point cloud and the grasping guide is selected. The robot is controlled to approach the object through the ROS system, and the grasping operation is completed by using a dexterous hand. During the grasping process, the grasping gesture is adjusted through tactile information to achieve stable grasping.

[0045] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dexterous grasping control of a mobile robot based on tactile feedback, the method comprising: Includes the following steps: The point cloud data of the target object is acquired, and multiple candidate grasping poses are generated using a conditional variational autoencoder. The candidate grasping poses include the position and orientation of the end effector and the dexterous hand joint configuration. The optimal grasping guide is selected from candidate grasping poses based on the physical bounding degree index, which includes grasping width coverage and grasping depth coverage. The optimal grasping guidance, robot body state, object relative pose, and tactile contact signals are input into a reinforcement learning strategy based on proximal strategy optimization to generate whole-body collaborative control commands to drive the robot to perform grasping, and the grasping gestures are adjusted in real time according to tactile signals during the grasping process.

2. The method for dexterous grasping control of a mobile robot based on haptic feedback according to claim 1, characterized in that, Including the following steps: Geometric features are extracted from the normalized object point cloud using a PointNet encoder. Latent variables are sampled from a Gaussian distribution and concatenated with geometric features before being input into the decoder to reconstruct the grasping pose; By sampling different latent variables multiple times, several candidate grasping postures with differences in spatial distribution and hand configuration are generated.

3. The method for dexterous grasping control of a mobile robot based on haptic feedback according to claim 1, characterized in that, Includes preliminary filtering and physical enclosure assessment: The initial filtering includes executability filtering and collision detection filtering. The executability filtering excludes non-forward-facing grasping based on the forward grasping space, while the collision detection filtering excludes candidate solutions where the end effector has descended below the lowest point of the object. The physical enclosure assessment calculates a comprehensive quality score based on the grasp width coverage and grasp depth coverage, and selects the grasp pose with the highest quality score as the optimal grasp guide.

4. The method for dexterous grasping control of a mobile robot based on haptic feedback according to claim 3, characterized in that, The calculation of the crawl width coverage includes: Define the vector connecting the thumb point and the finger point as the finger grip axis, and calculate the projection interval of the object's point cloud along this axis. The ratio of the intersection length of the finger's grasping area and the object's projection area to the total length of the object's projection area is calculated and used as the grasping width coverage of that finger.

5. The method for dexterous grasping control of a mobile robot based on haptic feedback according to claim 4, characterized in that, The calculation of crawl depth coverage includes: Define the projection from the center of the palm to the finger position in the direction perpendicular to the palm as the grip depth interval, and calculate the projection interval of the object point cloud in this depth direction. The ratio of the intersection length of the grip depth range of the finger and the palm with the object projection range to the total length of the object projection range is used as the grip depth coverage of that finger.

6. The method for dexterous grasping control of a mobile robot based on haptic feedback according to claim 1, characterized in that, The formula for calculating the overall quality score is: in, Let the grasping width coverage of the k-th finger be . Let k be the grasping depth coverage of the k-th finger. This represents the gripping depth coverage of the thumb.

7. The method for dexterous grasping control of a mobile robot based on haptic feedback according to claim 1, characterized in that, The input observation vectors for a reinforcement learning strategy include: The linear and angular velocities of the chassis, the angles of each joint of the robotic arm, and the angles of each joint of the dexterous hand constitute the body state; The relative pose information is formed by the three-dimensional coordinates of the object in the robot coordinate system and its relative distance to the dexterous hand; Binary contact signals are transmitted by tactile sensors installed in the fingertips, finger pads, and palm areas; And optimal crawling guidance information.

8. The method for dexterous grasping control of a mobile robot based on haptic feedback according to claim 1, characterized in that, Real-time adjustment of grip gestures based on tactile signals includes: When the tactile sensor on a finger does not detect a touch, the finger joint where the sensor is installed automatically tightens to form a tighter contact; When an object is detected to be sliding, the finger joint torque is dynamically adjusted to maintain a stable grip.

9. The method for dexterous grasping control of a mobile robot based on haptic feedback according to claim 1, characterized in that, Including the setting of phased reward functions: At a distance of 0.5 meters from the object, the robot is encouraged to approach the target at high speed and maintain the height of the dexterous hand by rewarding the arm's operating radius, chassis movement speed, chassis movement direction, and height. Within 0.1 meters of the object, the robot is encouraged to precisely align and close its dexterous hand to form multi-point contact through grasping posture alignment rewards, dexterous hand movement rewards, stable grasping rewards, and tactile rewards.

10. A dexterous grasping control system for a mobile robot based on haptic feedback, characterized in that, The tactile feedback-based dexterous grasping control method for mobile robots according to any one of claims 1-9 is applicable.

Images