Multi-Objective Deep Reinforcement Learning for Dimensional Optimization of Parallel Robots

Abstract

Dimensional parameter optimization of parallel robots faces two efficiency challenges: computationally intensive workspace sampling and the absence of reusable models across changing design requirements. This letter proposes a dual-layer framework. At the evaluation layer, Prior-Radius-Based Boundary Determination (PRBD) and Point-Array Boundary Determination (PABD) provide alternative boundary-focused workspace evaluation methods: PRBD prioritizes speed, while PABD prioritizes boundary accuracy. At the optimization layer, a Multi-Objective Deep Reinforcement Learning (MODRL) algorithm learns reusable policies using discrete actions, attention-enhanced LSTM networks, and a constraint satisfaction reward mechanism. Experiments on a 6-PUS parallel robot show that the proposed boundary-focused methods reduce sampling compared with polar coordinate traversal. C-indicator analysis shows stronger dominance than evolutionary and continuous-action baselines. Constraint satisfaction reward halves convergence time, and discrete actions outperform continuous alternatives. Transferability—the pre-trained policy reduces optimization time by 19.6% for new tasks while achieving a competitive Pareto front. Physical prototypes confirm that solution re-selection addresses design changes without re-optimization.