An RL-Driven Framework for Type-Dimension Co-Synthesis of Robotic Mechanisms

The type-dimension co-synthesis of robotic mechanisms presents a fundamental challenge due to the combinatorial complexity of topological architectures and the high-dimensional, nonlinear coupling between geometry and motion performance. This paper introduces a reinforcement learning (RL) framework that integrates graph grammar with graph neural networks (GNNs) to automate the co-synthesis process and can handle closed-loop chain mechanisms effectively. Our methodology establishes a systematic three-phase workflow: (1) constructs an comprehensive and diverse design space through graph grammar, enabling efficient mechanism synthesis via greedy algorithms guided by GNN-based performance prediction, (2) implements a rigorous multi-fidelity evaluation incorporating degree-of-freedom (DOF) verification, local transmission index (LTI) analysis, manufacturability constraints, and workspace assessment to ensure physically realizable designs, and (3) employs closed-loop optimization that continuously improves topological generation through GNN training. The framework is validated through comprehensive case studies on both parallel mechanisms (PMs) with 2 rotational and 1 translational (2R1T) and serial (6-DOF) manipulators, demonstrating superior performance compared to other search baselines.