Accelerating Vision-Language-Action Model Integrated with Action Chunking via Parallel Decoding

The performance of VLA models can be improved by integrating with # Action Chunking for effective control. However, action chunking linearly scales up action dimensions in VLA models with increased chunking sizes. This paper aims to tackle this problem by proposing PD-VLA, a parallel decoding framework for VLA models. This paper tries to use Jacobi Decoding method in VLA model.