Learning Dynamic Tasks on a Large-scale Soft Robot in a Handful of Trials

TL;DR

This work proposes a data-efficient Bayesian optimization-based approach for learning control policies for dynamic tasks on a large-scale soft robot that optimizes the task objective function directly from commanded pressures, without requiring approximate kinematics or dynamics as an intermediate step.

Abstract

Soft robots offer more flexibility, compliance, and adaptability than traditional rigid robots. They are also typically lighter and cheaper to manufacture. However, their use in real-world applications is limited due to modeling challenges and difficulties in integrating effective proprioceptive sensors. Large-scale soft robots ($\approx$ two meters in length) have greater modeling complexity due to increased inertia and related effects of gravity. Common efforts to ease these modeling difficulties such as assuming simple kinematic and dynamics models also limit the general capabilities of soft robots and are not applicable in tasks requiring fast, dynamic motion like throwing and hammering. To overcome these challenges, we propose a data-efficient Bayesian optimization-based approach for learning control policies for dynamic tasks on a large-scale soft robot. Our approach optimizes the task objective function directly from commanded pressures, without requiring approximate kinematics or dynamics as an intermediate step. We demonstrate the effectiveness of our approach through both simulated and real-world experiments.

Figures (5)

• Figure 1: The soft robot design used in this paper for simulation and hardware-based experiments. This is a 12 degree of freedom (DOF) pneumatically actuated, large-scale soft robot about 1.3 meters in length. It consists of three continuum joints connected by two rigid links.
• Figure 2: An illustration of the simulated version of the physical robot. Each continuum joint is approximated by individual disks that are connected by universal joints with springs and dampers connected in parallel, shown in green. Compressive forces from pneumatic actuators are applied between disks, along the $x$ and $y$ axes, shown in red and blue, respectively.
• Figure 3: Throwing task: The robot starts with the gripper holding the object \ref{['fig:sim-throw:a']}), generates high end-effector velocity, and has to release the object the right time \ref{['fig:sim-throw:b']}) to throw successfully.
• Figure 4: Hammering task. Starting from the equilibrium state (\ref{['fig:hardware']}), the robot has to "wind-up" enough energy \ref{['fig:sim-hammer:a']}) to activate the touch force sensor \ref{['fig:sim-hammer:b']}).
• Figure 5: Tip velocities that resulted from using BayesOpt-LEI, BayesOpt-UCB, and a random policy.