A Conformal Mapping-based Framework for Robot-to-Robot and Sim-to-Real Transfer Learning

TL;DR

This paper introduces a conformal mapping-based framework that enables effective transfer of motion planning and control policies between robots and from simulation to real-world, addressing system differences and aging effects.

Contribution

It proposes a novel Schwarz-Christoffel mapping method for geometric transfer of control inputs and a primitive motion generation approach for compatibility with different robot capabilities.

Findings

Successful transfer of policies in simulations and real robots

Reduction of sim-to-real transfer gap

Robustness to system aging and failures

Abstract

This paper presents a novel method for transferring motion planning and control policies between a teacher and a learner robot. With this work, we propose to reduce the sim-to-real gap, transfer knowledge designed for a specific system into a different robot, and compensate for system aging and failures. To solve this problem we introduce a Schwarz-Christoffel mapping-based method to geometrically stretch and fit the control inputs from the teacher into the learner command space. We also propose a method based on primitive motion generation to create motion plans and control inputs compatible with the learner's capabilities. Our approach is validated with simulations and experiments with different robotic systems navigating occluding environments.