TE-RoboNet: Transfer Enhanced RoboNet for Sample-Efficient Generation of Robot Co-Designs

Kishan Reddy Nagiredla, Arun Kumar A V, Kevin Sebastian Luck, Thommen George Karimpanal, Santu Rana

Keywords: Co-Design, Reinforcement Learning, Robot Design

Abstract: Robot co-design requires joint optimization of morphology and its control mechanism and is thus associated with a vast, high-dimensional design space. Traditional co-design methods are sample-inefficient, and are thus used for incremental refinement of known designs rather than the discovery of novel, high-performing embodiments by effectively traversing this complex space. Our method extends RoboNet, a novel Generative Flow Network based robot co-design method that excels in generating superior designs in a sample-efficient manner but does that with independent training of control mechanism for each individual robot, resulting in good designs paired with weak controllers. In this paper, we propose a novel policy transfer mechanism to continuously learn a modularized policy, comprising a core network shared across all robot morphologies, and morphology-specific adapters. By effectively disentangling morphology-specific and transferable control components, our framework addresses the critical challenge of knowledge transfer between robot morphologies and their topologies with varying DoFs. Experiments in four distinct co-design environments show that our method, TE-RoboNet, achieves up to 40% improvement in performance compared to the closest co-design baselines under equivalent memory and computational budgets.

Confirmation: I understand that authors of each paper submitted to EWRL may be asked to review 2-3 other submissions to EWRL.

Serve As Reviewer: ~Kishan_Reddy_Nagiredla1

Track: Regular Track: unpublished work

Submission Number: 171