AI

Gated VLA-Cache Recomputes Only When the Policy Sounds Unsure

Robb Harlan 4 min read

VLA inference is expensive because the same 7B transformer re-encodes a camera frame that barely moved. VLA-Cache already reuses key-value states for static patches. University of Tokyo researchers (Zhijie Wu, Kento Kawaharazuka, Kei Okada) add a second check: if the model’s top-two action tokens are too close, throw the cache away. The paper (arXiv:2608.10824, August 11, accepted IROS 2026) calls it Gated VLA-Cache.

Split view of a robot arm in VLA-Cache mode versus full recompute after uncertainty detection
When the logit margin collapses, the gate forces a full visual recompute. Source: Wu, Kawaharazuka, and Okada, arXiv:2608.10824.

How the gate works

VLA-Cache marks patches with cosine similarity above 0.996 as static, then filters out ones that still get high text-to-vision attention, then reuses more tokens in low-entropy layers. That is all observation-space. Gated VLA-Cache averages the top-1 minus top-2 probability across the action tokens from the previous step. If that margin is below θm (0.65 for OpenVLA, 0.50 for OpenVLA-OFT), the entire KV cache is invalidated.

No extra parameters. No retraining. The signal is already sitting in the softmax.

LIBERO, two architectures

Four suites, 50 episodes × 10 tasks, NVIDIA A100. OpenVLA:

SuiteFullVLA-CacheGated
Spatial78.8% / 1.89 TFLOP78.8% / 1.4379.4% (paper text)
Goal77.2% / 1.8374.0% / 1.40recovers 106% of the 3.2-point drop (77.4%)
Long54.0% / 1.8850.2% / 1.4354.8% at 1.54 TFLOP (121% of the lost gap)
Average70.1% / 1.8768.1% / 1.4369.9% / 1.53

That average is 18% cheaper than full inference and 0.2 points off full accuracy. On LIBERO-Long the gate fires on about 24% of steps, clustered on grasp alignment. On OpenVLA-OFT, where blind caching barely hurts (95.8% either way), the gate stays quiet.

Four LIBERO suite scenes: Spatial, Object, Goal, and Long
LIBERO suites used for the OpenVLA and OFT sweeps. Source: Wu et al., arXiv:2608.10824.

Thresholds are stable in a plateau (0.55–0.75 on OpenVLA Spatial stays 78.2–79.4%). The authors say they have not run this on a physical robot yet.

A Human’s Take

Recompute when the policy is arguing with itself is a grown-up cache. I like that it is training-free and that it mostly sleeps when OFT is already calm. The one-step lag is real: yesterday’s doubt decides today’s pixels. Until someone times this on a wrist camera at 15 Hz, it is a LIBERO result. Still, if you already ship VLA-Cache, this gate is the cheap insurance I would turn on.

Sources