Experimental Investigation of Simplified Grasping and Testing Strategies for a Hand-Like Robotic Gripper

Sammanfattning

This project evaluates and compares the performance of an underactuated anthropomorphicrobotic hand with conventional two-finger grippers in pick-and-place and object discrimination tasks using only proprioceptive sensory data. The project aims to simplify sensory implementations for robotic grasping and object discrimination, since these tasks frequently require external sensors, such as force or visual sensors.The experimental setup consisted of a Universal Robots UR5 on which the anthropomorphic hand, the qb SoftHand Research, and the OnRobot RG6 grippers were mounted. Also, a Universal Robot UR10e with a Robotiq 2F-85 was needed due to hardware failures. The pick-and-place and object discrimination comparisons were done under controlled conditions, using two different sets of objects for each part. The qualitative metrics used for the pick-and-place task were repeatability, success rate and slip ratio. For the object discrimination task, a classification model based only on internal sensory feedback was implemented, using motor current and the actual position. The supervised machine learning algorithms Support Vector Machine (SVM) and RandomForest (RF) were trained but only validated for the Robotiq 2F-85 gripper due to the anthropomorphic hand’s failure during testing.The obtained results and their subsequent analysis show the behavioural differences between the grippers in pick-and-place tasks. Furthermore, they demonstrate that object discrimination using only internal data is feasible, under controlled conditions and with a reduced object set, highlighting limitations related to overfitting or the dataset size.