Behavior Cloning from Observations with Domain Mapping for the Control of Soft Robots

Abstract

The designs of soft robots often draw inspiration from biological systems, even though their morphologies and actuation mechanisms often differ significantly. Reproducing bio-inspired movements on a corresponding platform neces-sitates in-depth knowledge of these mechanisms. This paper introduces a simplified approach to tackle this challenge in an imitation learning-based problem setting. Leveraging a human arm as an expert, state-only paths were recorded and quantitatively replicated on a pneumatically actuated two-module soft arm. To overcome morphological differences, both bodies were segmented into a fixed number of segments. Various mapping strategies were explored to project the observed workspace of each segment from the demonstrator's body onto the corresponding operational workspace of the learner. A platform-specific model was trained to predict actions for reaching the mapped expert demonstrations. Ultimately, a policy trained from these state-action pairs using behavioral cloning successfully imitated the expert demonstrations, achieving a mean MAE consistently under 3% of the overall workspace of the soft robotic arm. On average, the proposed algorithm achieved above 79% mean MAE reduction in 16 episodes, with each episode approximately 7 sec long, across three different tasks.