Editable progressive generative adversarial networks for locomotion skill learning in space quadruped robots

Locomotion skills are fundamental for space robots to perform diverse space missions, such as on-orbit maintenance, large spacecraft assembly, and extraterrestrial exploration, requiring high autonomy and environmental generalization capability. However, mainstream methods under current reinforcement learning and imitation learning paradigms face significant challenges, including insufficient generalization and reliability, when applied to adhesive locomotion skill learning for space robots. To address these issues, this paper introduces generative adversarial algorithms into the imitation learning paradigm, proposing an editable progressive generative adversarial network (EPGAN). A regulator module is incorporated into the conventional generator-discriminator architecture. The regulator optimizes the trajectories output by the generator and provides explicit convergence guidance to the generator via an L2 loss. Concurrently, it probabilistically replaces the original generated trajectories with the optimized ones as negative samples for the discriminator. This strategy increases the difficulty of discriminator training, effectively balances the adversarial process, mitigates generator gradient vanishing, and significantly enhances the model’s learning capability on low-redundancy, noise-sensitive trajectory data. An editing network is embedded between the trained generator and the latent variables. Guided by task scenario attributes and motion constraint knowledge, it performs generative adversarial inverse inference via gradient descent to generate reliable locomotion trajectories satisfying specific environmental boundary constraints. Experimental validation demonstrates that in three test tasks involving different gaits, EPGAN achieves a success rate exceeding 81% under a 5% trajectory error tolerance. In more challenging generalization tasks, it maintains a success rate superior to 61% under a stricter 1% error tolerance, significantly outperforming existing comparative methods.