Lune

CVPR Findings2026

Retrieval-VLA: Training-Free In-Context Adaptation for Vision-Language-Action Models

Yue Zhang, Rui Wang, Jiehong Lin, Zhongrui Wang, Xiaojuan Qi

Abstract

Embodied AI aims to build agents that can perceive, reason, and act in the world. While Vision-Language-Action (VLA) models have made significant strides towards this goal by learning from large-scale datasets, they remain fundamentally static. Their knowledge, encoded within fixed weights, fails to adapt rapidly to new objects, environments, or instructions without costly fine-tuning. In contrast, humans excel at in-context adaptation by retrieving and reusing relevant past experiences. Inspired by this capability, we introduce Retrieval-VLA, a novel framework that augments pretrained VLAs with a training-free, plug-and-play retrieval mechanism to enable dynamic in-context adaptation. Our approach abstracts scenes into sparse, object-centric keypoints, discarding irrelevant visual clutter while preserving critical affordances. During inference, a lightweight cross-attention module uses the current scene as a query to retrieve the most relevant action plans from a memory buffer of past demonstrations, which are then fused with the base VLA's output. This process requires no parameter updates to the underlying model. Evaluations on LIBERO and CALVIN benchmarks show consistent improvements in zero-shot and few-shot performance over π 0 and Open-VLA backbones, demonstrating the practical viability of training-free retrieval for generalist, adaptive, and dataefficient embodied intelligence.

Ask Lune

Know this paper in minutes, not an afternoon.

Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.

Questions to start from

Lune can draw on

the full paper · related work

Get the answer

Free to start. No credit card required.

Lineage

  • TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies cites Visual Instruction Tuning
  • Interleave-VLA: Enhancing Robot Manipulation with Image-Text Interleaved Instructions cites Visual Instruction Tuning
  • MemoryVLA: Perceptual-Cognitive Memory in Vision-Language-Action Models for Robotic Manipulation cites TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies
  • MemoryVLA: Perceptual-Cognitive Memory in Vision-Language-Action Models for Robotic Manipulation cites Visual Instruction Tuning
Builds onCited byLeft to right by year; the more cited, the higher.

Builds on6

  1. Visual Instruction TuningHaotian Liu, Chunyuan Li, Qingyang Wu, Yong Jae LeeNeurIPS 2023 · 11,349 citations
  2. MemoryVLA: Perceptual-Cognitive Memory in Vision-Language-Action Models for Robotic ManipulationHao Shi, Bin Xie, Yingfei Liu, Lin Sun et al.ICLR 2026 · 227 citations
  3. Interleave-VLA: Enhancing Robot Manipulation with Image-Text Interleaved InstructionsCunxin Fan, Xiaosong Jia, Yihang Sun, Yixiao Wang et al.ICLR 2026 · 49 citations
  4. VISA: Retrieval Augmented Generation with Visual Source AttributionXueguang Ma, Shengyao Zhuang, Bevan Koopman, Guido Zuccon et al.ACL 2025 · 24 citations
  5. TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic PoliciesRuijie Zheng, Yongyuan Liang, Shuaiyi Huang, Jianfeng Gao et al.ICLR 2025

Related papers

  1. RA-VLA: Retrieval-Augmented VLA for Test-Time AdaptationSanghwan Jang, Minjin Jeon, Minsoo Kim, Seong Jin Choi et al.ICML 2026
  2. MetaVLA: Unified Meta Co-Training for Efficient Embodied AdaptationChen Li, Zhantao Yang, Han Zhang, Fangyi Chen et al.ICLR 2026 · 2 citations
  3. A1: Adaptive Truncated Vision-Language-Action Model from Affordance to ActionKaidong Zhang, Jian Zhang, Rongtao Xu, Yu Sun, Youpeng Wen, Shuoshuo Xue, Xiaoyu Guo, Minghao Guo, Weijia Liufu, Liu Zihou, Kangyi Ji, Zihang Li, Ruiyi Chen, Meng Cao, Jingming Zhang, Shen Zhao, Xiaojun Chang, Feng Zheng, Ivan Laptev, Xiaodan LiangCVPR Findings 2026
  4. Think Less, Act Early: Reinforced Latent Reasoning with Early Exit in Vision-Language-Action ModelsDianqiao Lei, Lianlei ShanICML 2026
  5. FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action AdaptationDuc Nguyen, Nghiem Diep, Binh Nguyen Gia, Trong-Bao Ho et al.ICML 2026 · 3 citations
Dusk over the sea between two cliffs drawn in fine vertical lines