A Smart Grasping System for Handling Irregular, Naturally Varying Objects

This paper presents the design, integration, and validation of a smart grasping system for handling irregular, naturally varying objects. The system consists of a 6-axis robot, a soft robotic gripper, a vision sensor and a computer. A grasping algorithm utilizing reinforcement learning is imple-mented to provide the flexibility and adaptiveness required to handle object variations. Benchmark testing were conducted on simple objects and the system achieved a 68 % grasp success rate after 1500 training iterations. Improvements to the system were then implemented including the repositioning of the vision sensor, a reset mechanism and a collision avoidance algorithm. A grasp success rate of 80% was achieved with the improved system. Kumara (sweet potato) was selected in this case as an example of irregular, naturally varying objects. Initial training and testing with kumara proved to be challenging and a pre-training approach with annotated images were proposed and implemented. Human grasping experience was incorporated into the grasping system via the pre-training and a 71 % grasp success rate was achieved after 1500 iterations.