VLA-Precision: Asymmetric Co-Bootstrapping for Efficient Real-World Online RL of Vision-Language-Action Models

To improve the training efficiency of the visual-language-action model in precision tasks, researchers proposed the VLA-Precision framework. This framework uses an algorithm based on Asymmetric Cooperative Guidance (ACoB) and achieves state decoupling and on-demand streaming processing through the ACoB-Stream architecture, effectively addressing the issues of unreliable value signals and high computational costs for large models. Experimental evaluations show that in four categories and nine high-precision chemical tasks, VLA-Precision achieved an average success rate of 98.3%, with a single-task duration of 45.8 minutes. Its running speed is 1.2 times and 1.8 times higher than that of the VLA baseline and pure reinforcement learning (RL) baseline, respectively, and its overall throughput increased by 10.9 times.