ARViP: Adversarial Regularization in Visuomotor Policies for Robotic VLA Purpose

With the development of embodied intelligence, current vision-language-action (VLA) models are difficult to deploy due to their large number of parameters thus high computational load, while the vision-based action models suffer from insufficient action accuracy in grasping tasks. To address these issues, this article proposes a novel Adversarial Regularization in Visuomotor Policy (ARViP) model. First, a proposed conditional variational autoencoder generative adversarial network architecture is applied to the existing generative action models. By discriminator distinguishing between real and generated samples, the generator is forced to produce increasingly realistic outputs, thereby increasing the accuracy of grasping. Second, to address the issue of causal confusion, we introduce a target detection model incorporating prior knowledge of grasped objects as the vision backbone. It allows the vision backbone to have prior knowledge of the feature space for grasped objects, which strengthens the object localization capabilities, and, in turn, improves vision understanding and processing capabilities. Finally, to address the issue of current models typically using uniform time-step optimization for all robotic arm joints, we propose a time-based joint regularization method. It introduces specialized loss functions tailored to the individual joint states at different time steps, thereby improving the prediction accuracy, smoothness, and robustness of motion sequences. In several real-world experiments, including assembling, disassembling, and stacking tasks, the proposed ARViP model improves grasping accuracy by 32% compared to the studied models.