Event
rMuscle uses two-stage caching to accelerate VLA inference by 1.29–1.42×
1 reports · 1 sources
Kaijun Zhou and other researchers propose rMuscle, a real-time vision-language-action (VLA) model inference framework inspired by human muscle memory. The paper's analysis finds that when a robot repeatedly performs tasks, there are significant similarities in observations, action trajectories, and even internal model states. rMuscle uses two-stage caching: Context Cache reuses visual token outputs to reduce computation, and Action Cache reuses neuron activation patterns to reduce weight accesses. On an RTX 4090 and Jetson Thor, for LIBERO, RoboTwin, and real-robot manipulation tasks, rMuscle achieves 1.29–1.42× speedup while maintaining the original success rate on real robots.
Report timeline
- rMuscle: A Real-Time Framework That Uses Robotic Task Similarity to Accelerate VLA Model Inference
The paper proposes rMuscle, a real-time vision-language-action (VLA) model inference framework inspired by human muscle memory. The authors first analyze embodied workloads and find that when robots repeatedly perform tasks, there is significant similarity in observations, action trajectories, and even internal model states.
arXiv cs.RO ·