Research map / Vision-language and multimodal / Robot action and navigation
VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Training-free VLA inference acceleration that caches KV entries of static visual tokens across frames, re-computes task-relevant tokens selected by decoder attention, and adapts reuse ratio per layer via attention entropy.
Problem
VLA models are computationally expensive for real-time robotic control, and existing acceleration methods either need retraining or ignore the temporal redundancy of closed-loop manipulation.
Approach
Reuse cached key-value entries of visually static image patches between adjacent frames, exclude task-relevant tokens (high text-to-vision attention) from reuse, and set per-layer reuse ratios from cumulative attention-entropy reduction.
Key results
- On LIBERO with OpenVLA, VLA-Cache reaches 74.7% average success vs 75.0% baseline while cutting latency from 51.91 ms to 31.83 ms and raising control frequency from 4.23 to 4.59 Hz (FLOPs 1.864T to 1.355T).
- On LIBERO with OpenVLA-OFT, VLA-Cache keeps average success at 97.4% (baseline 96.8%) and raises control frequency from 65.10 to 78.98 Hz.
- On SIMPLER with CogACT, VLA-Cache keeps success within about 1 point of baseline (74.4% vs 74.8% matching; 62.3% vs 61.3% aggregation) and raises control frequency from about 12.4 to about 14.5 Hz.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply VLA-Cache to OpenVLA | OpenVLA · LIBERO | latency: 51.91 -> 31.83 | Helped |
| Apply VLA-Cache to OpenVLA | OpenVLA · LIBERO Average | success rate: 75.0% -> 74.7% | No clear effect |
| Apply VLA-Cache | OpenVLA-OFT · LIBERO | control frequency: 65.10 -> 78.98 | Helped |
| Apply VLA-Cache | CogACT · SIMPLER Visual Matching | latency: 54.29 -> 39.63 | Helped |
| Apply VLA-Cache | CogACT · SIMPLER Visual Matching Average | success rate: 74.8% -> 74.4% | No clear effect |
| Apply VLA-Cache | CogACT · SIMPLER Variant Aggregation Average | success rate: 61.3% -> 62.3% | Helped |
| Apply VLA-Cache | OpenVLA · Real robot Average | success rate: 82.1% -> 84.6% | Helped |
| Reuse visually static tokens without filtering task-relevant tokens (ablation) | OpenVLA · LIBERO Spatial | success rate: 84.4 -> 74.2 | Hurt |
| Add task-relevance filtering to static-token reuse (ablation) | OpenVLA · LIBERO Spatial | success rate: 74.2% -> 82.6% | Helped |
| Add layer-adaptive token reuse (ablation) | OpenVLA · LIBERO Spatial | success rate: 82.6% -> 83.8% | Helped |
| Replace attention-based task relevance with object masks (ablation) | OpenVLA-OFT · LIBERO-Spatial | success rate: 98.3 -> 87.4 | Hurt |
| Increase static-token budget across tested settings (ablation) | OpenVLA-OFT with VLA-Cache · LIBERO-Spatial | success rate: 97.6; 97.8; 98.2; 98.0; 97.4; 96.6 | Mixed |
Surprising findings
- Naively reusing all visually static tokens drops OpenVLA LIBERO-Spatial success from 84.4% to 74.2%, showing pixel similarity alone is insufficient.
- FastV and SparseVLM fail to improve speed on VLA models and often degrade success, since VLA decoders generate short action outputs.
Limitations
- Acceleration decreases in dynamic scenes with large background or object motion, since more tokens must be recomputed.
- Evaluated only on LLaMA2-based VLA decoders; applicability to other backbones such as Gemma-based pi0 is untested.
- Real-world evaluation is small (100 trials total) and uses a single baseline; the reported gains are within a few percentage points.
Experimental setup
Datasets: LIBERO SIMPLER CLVR_Jaco_Play
Benchmarks: LIBERO SIMPLER real-world Kinova Jaco2 tasks
Metrics: success rate FLOPs CUDA latency control frequency
Baselines: OpenVLA SparseVLM FastV CogACT OpenVLA-OFT
Builds on: FastV KV caching LoRA
Closest papers by meaning
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.94)
- VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.93)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models (CVPR 2025 · similarity 0.92)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.92)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.