COGENT: Co-design of Robots with GFlowNets

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

Keywords: Robot Design, Reinforcement Learning, Embodied Intelligence, Robot Co-Design

Abstract: Co-design of robots involves optimizing the control mechanism and physical form together. This intertwined design process is inherently challenging and sample inefficient because of the large design and control search spaces. We introduce COGENT, a novel framework that leverages a graph synthesis technique named GFlowNet, to enhance search space traversal in robotic co-design. To increase sample efficiency, the proposed framework introduces a cost/performance-aware design prioritization mechanism that learns a design generator policy by carefully sampling the design space. Our experiments show the effectiveness of the proposed framework in various robot co-design tasks. Evaluations performed on a wide range of agent design problems demonstrate that our method significantly outperforms baselines. We show that COGENT produces a suite of diverse designs achieving better task objectives across all evaluated design problems.

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Track: Regular Track: unpublished work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: 151