DepthCache: Depth-Guided Training-Free Visual Token Merging for Vision-Language-Action Model Inference