Optimizing Robot Control Through Action Sensitive Compression

Summary & Methodology Analysis

Vision-language-action models often face significant latency during closed-loop control because existing compression techniques rely on generic visual similarity metrics. These proxies ignore how specific token changes propagate into physical action errors. To address this, the researchers developed Action-JND, which trains a lightweight estimator on a frozen policy to predict feature perturbations. The system identifies the maximum allowed deviation for tokens that still keeps the resulting robot action within a set tolerance threshold. This process generates action-tolerance scores that rank tokens by their physical importance, allowing the system to selectively prune or reuse cache states without degrading task reliability. The architecture leverages these scores within standard compression pipelines to prioritize which data to compress during inference. By integrating this ranking signal into KV-cache reuse and token pruning workflows, the system ensures that critical information governing the robot movement remains intact while non-essential tokens are minimized. The researchers validated the approach using OpenVLA and OpenVLA-OFT models on the LIBERO benchmark. This systematic focus on the relationship between latent tokens and physical actuator output provides a more robust framework for deployment in resource-constrained environments. While the results demonstrate improved reliability, the current evaluation is limited to single-view and dual-view configurations. The authors note that the method requires further extension to accommodate more diverse embodiments and complex action spaces in future applications.