Just Noticeable Difference Modeling for Token Compression in Vision-Language-Action Models

TL;DR - Action-JND estimates how much each visual token in a vision-language-action policy can change without exceeding a tolerated action deviation. This enables safer token pruning and stale-KV reuse for lower-latency robotic control, particularly at aggressive compression ratios.

• Defines token “noticeability” using language-conditioned action responses rather than indirect similarity, attention, or saliency signals.
• Uses a lightweight token-wise estimator in deep visual-feature space to predict maximum action-preserving perturbations.
• Produces action-tolerance scores that plug into multiple compression methods and prioritize tolerant tokens for compression.
• Experiments with OpenVLA and OpenVLA-OFT on LIBERO report consistently improved compression reliability.