Comparing Model Based and Model Free techniques for Underactuated In-hand Manipulation

Abstract

Compared to their fully actuated counterparts underactuated hands are cheap, lighter and provide stable grasp across variety of objects without feedback. However, underactuated hands are less dexterous for in-hand manipulation task due to the limited range of motion in their configuration space. Brake Assisted Tendon Actuator (BATA) is a novel mechanism to enhance dexterity in underactuated hand. This work aims to implement a controller framework for BATA and asses it’s in hand manipulation capabilties. Control is challenging due to contacts, under-actuation and model uncertainty. We develop a simulation environment in MuJoCo and formulate the underlying discrete Markov Decision Process. Model based and model free reinforcement learning methods are implemented to learn a policy for a specific type of in-hand manipulation task: rolling. Simulation results with objects of varying mass and radius suggest Model Predictive Path Integral (MPPI) is more generalizable compared to model free, Proximal Policy Optimization (PPO).