Go to 2024 Workshop MAPoDeL homepageRoboNet: A Sample-Efficient Robot Co-Design Generator
Keywords: Co-Design, Robotics, Machine Learning
Abstract: Co-design of robots involves designing 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. Our key idea is to navigate this combinatorial search space more intelligently than earlier methods to reduce dependence on excessive samples while still generating capable, complex designs. Our proposed framework RoboNet, leverages a GFlowNet-based approach, a recent advancement in graph synthesis to explore the vast, high-dimensional space of robot co-designs efficiently. RoboNet learns to generate robots from scratch and evaluates designs based only on partially trained control policies. Specifically, we use GFlowNet in combination with a rate-based design prioritizing and cost-aware sampling strategy to evaluate robots based on the future promise under full-training and across different complexities. Our experiments show the proposed framework's utility in various robot design tasks.
Submission Number: 3
Loading