Knowledge Bank

The dexterity of the human hand is often taken for granted. Even seemingly simple tasks to humans like moving an object around in the hand can be complex and require coordination from many different muscles. For this reason, robots have a difficult time trying to match the in-hand manipulation capabilities of humans. For robotic hands to become helpful tools in the fabrication, manufacturing, and packaging industries, more advanced control strategies are needed. In this research, two strategies are investigated to attempt to improve the training of a reinforcement learning controller during simple manipulation tasks. This project was done in simulation using MATLAB Simscape to study two fingers manipulating a small box. The controller for the system was created using reinforcement learning agents from MATLAB's Reinforcement Learning Toolbox to learn a controller for manipulating the box. The results from this study showed an improvement in the policy learned when the initial conditions of the trainings were randomized. This result also showed the ability for a learned policy to generalize further than just the conditions used for training and complete the same task in new environments. We also explored the effect of under-actuated versus fully actuated fingers in manipulating a box.