rMuscle: Robotic Muscle Memory for Efficient Vision-Language-Action Model Inference

TL;DR - rMuscle accelerates Vision-Language-Action model inference by caching reusable visual-token outputs and neuron activation patterns across repetitive robot executions. It improves responsiveness by 1.29–1.42× without reducing real-world task success rates.

• A Context Cache reuses visual-token outputs to avoid redundant computation.
• An Action Cache reuses activation patterns to reduce model-weight accesses.
• Online recomputation, sliding-window retrieval, and shared masks limit cache memory and access overhead.
• Evaluations span RTX 4090 and Jetson Thor hardware, LIBERO, RoboTwin, and physical manipulation tasks.