Abstract: With advancements in science and technology, collaborative and industrial robotic arms are increasingly gaining popularity. Enhancing the intelligence and autonomy of robot arms, particularly in autonomous grasping, has become one of the research hotspots in robotics research. To improve the efficiency and success rate of industrial robot arms in grasping target objects and avoiding obstacles, a three-finger pneumatic flexible clamp was selected, and a flexible grasping module was designed. Communication between the upper computer and the single-chip computer via a serial port enables clamping and loosening actions, constructing an autonomous grasping system based on the traditional Informed -RRT^* algorithm. An improved info-RRT ^* algorithm (Grasping informed-RRT ^*, GI-RRT^*) for the GR-ConvNet model is proposed. First, the maximum number of iterations and the adaptive function are pre-set to shorten the generation time of the manipulator’s motion trajectory and enhance sampling guidance and quality. Second, direct sampling of elliptical subsets constrains the position of sampling points, improving sampling efficiency. Finally, a greedy algorithm deletes redundant path points, and a cubic B-spline curve smoothly constrains the trajectory of the robot arm, shortening its length and improving flexibility. The generated residual convolutional neural network (GR-ConvNet) model predicts inputs from color and depth images captured by a depth camera, outputting the appropriate mapping grab pose of the object in the field of view. To verify the grasping effect of the robot arm, simulation and grasping experiments were conducted on the cooperative robot arm FR3. Simulation results show that, compared with the traditional Informed-RRT^* algorithm, the improved algorithm shortens trajectory length by 10.11% and reduces trajectory generation time by 62.68%. The robot arm independently avoids obstacles and grasps target objects, meeting the requirements for autonomous grasping. Experiments with the cooperative robot arm demonstrate its ability to independently grasp objects independently and successfully avoid obstacles. This further validates the algorithm’s effectiveness on a real robot arm, bringing hope for its further development and use. It reduces the difficulty for operators to use the robot arm and accelerates the wide application of domestic robot arms in factories. This paper aims to promote the practical application of robot arms.