A Survey on Efficient Vision-Language-Action Models

Zhaoshu Yu; Bo Wang; Pengpeng Zeng; Haonan Zhang; Ji Zhang; Zheng Wang; Lianli Gao; Jingkuan Song; Nicu Sebe; Heng Tao Shen

Source-linked AI summary

A Survey on Efficient Vision-Language-Action Models

Zhaoshu Yu, Bo Wang, Pengpeng Zeng, Haonan Zhang, Ji Zhang, Zheng Wang, Lianli Gao, Jingkuan Song, Nicu Sebe, Heng Tao Shen

arXiv:2510.24795v2Oct 27, 2025cs.CVcs.AIcs.LGcs.RO

TL;DR

Foundational VLAs face computational, latency, and data burdens that impede real-time deployment on physical platforms, while existing efficiency research lacks a unified review. This survey synthesizes the field across the model-training-data pipeline using a three-pillar taxonomy, and identifies trade-offs, challenges, and future directions for efficient embodied intelligence.

Problem
Foundational VLAs inherit prohibitive computational and data footprints from large-scale foundations, creating latency and power mismatches with physical platforms.
Method
The survey systematically organizes efficient VLA research into Efficient Model Design, Efficient Training, and Efficient Data Collection.
Results
The synthesis consolidates fragmented efficiency efforts and identifies interdependent pillars, trade-offs, challenges, and a roadmap toward adaptive, co-designed systems.
Takeaways & Limitations
The framework serves as a reference for navigating efficient VLA research and supporting scalable embodied intelligence.
Takeaways & Limitations
Efficient approaches face trade-offs between model compactness and multimodal expressivity, alongside alignment, reward, and distribution-shift challenges.

Abstract

from arXiv · show

Vision-Language-Action models (VLAs) represent a significant frontier in embodied intelligence, aiming to bridge digital knowledge with physical-world interaction. Despite their remarkable performance, foundational VLAs are hindered by the prohibitive computational and data demands inherent to their large-scale architectures. While a surge of recent research has focused on enhancing VLA efficiency, the field lacks a unified framework to consolidate these disparate advancements. To bridge this gap, this survey presents the first comprehensive review of Efficient Vision-Language-Action models (Efficient VLAs) across the entire model-training-data pipeline. Specifically, we introduce a unified taxonomy to systematically organize the disparate efforts in this domain, categorizing current techniques into three core pillars: (1) Efficient Model Design, focusing on efficient architectures and model compression; (2) Efficient Training, which reduces computational burdens during model learning; and (3) Efficient Data Collection, which addresses the bottlenecks in acquiring and utilizing robotic data. Through a critical review of state-of-the-art methods within this framework, this survey not only establishes a foundational reference for the community but also summarizes representative applications, delineates key challenges, and charts a roadmap for future research. We maintain a continuously updated project page to track our latest developments: https://evla-survey.github.io/.

1 INTRODUCTION

Foundational VLAs offer powerful multimodal reasoning and action but face latency, computational, and data bottlenecks that hinder real-time edge deployment. This survey addresses the gap with a three-pillar framework covering efficient model design, training, and data collection.

1 INTRODUCTION: Foundational VLAs inherit large computational and data footprints, creating latency and power mismatches with physical platforms.These constraints hinder transition from digital intelligence to real-time actuation on edge devices.
1 INTRODUCTION: Efficient VLAs target high-frequency control and robust performance across robotic scenarios while bridging heavyweight foundations and edge deployment.The paradigm optimizes efficiency across the model-training-data lifecycle.
1 INTRODUCTION: The survey organizes efficient VLA research into Efficient Model Design, Efficient Training, and Efficient Data Collection.The taxonomy covers architectural and inference efficiency, training-resource reduction, and scalable data acquisition and enrichment.
1 INTRODUCTION: The survey presents the first comprehensive review of efficient VLAs across the entire model-training-data process.It aims to establish a foundational reference for the research community.
1 INTRODUCTION: The survey critically distills field challenges and limitations and outlines future directions for scalable embodied intelligence.It is intended to help researchers and practitioners navigate the literature and inspire further research.

2 VISION-LANGUAGE-ACTION MODELS

VLAs combine visual encoding, multimodal language reasoning, and action decoding into an end-to-end pipeline for embodied manipulation, supported by diverse datasets and benchmarks. Their deployment is constrained by architectural, training, and data inefficiencies that motivate efficient VLA research.

2.1.1 Foundational Pipeline: VLAs integrate vision encoders, LLM backbones, and action decoders to map visual and linguistic inputs to robot control trajectories.Action decoders may be diffusion- or flow-based, autoregressive, or MLP-based.
2.1.2 Datasets: The VLA ecosystem relies on high-fidelity datasets and standardized benchmarks spanning diverse tasks, embodiments, environments, and manipulation scenarios.OXE aggregates multi-lab robotic data, while RLBench, RoboTwin, and Meta-World provide predefined task libraries and consistent evaluation metrics.
2.2 The Need for Efficient VLAs: Foundational VLAs face deployment barriers because massive transformer-based backbones cause prohibitive inference latency and inadequate control frequencies for dynamic robotic actuation.The survey identifies model architecture as a central source of inefficiency and presents parameter, latency, and operating-frequency comparisons in Table 1.
2.2 The Need for Efficient VLAs: Large-scale VLA pre-training demands substantial computation and energy, restricting development and generalization across heterogeneous embodiments to well-resourced institutions.The burden is associated with the computational requirements of large-scale pre-training.
2.2 The Need for Efficient VLAs: High-fidelity trajectory acquisition through real-world demonstrations or expert-designed simulation is time-consuming, difficult to scale, and costly.The data pipeline therefore forms a second major source of deployment inefficiency alongside model architecture.
2.2 The Need for Efficient VLAs: These inefficiencies create deployment bottlenecks in resource-constrained settings and motivate Efficient VLAs as a route toward practical robotic use.The survey frames efficiency as a prerequisite for broader deployment across applications such as autonomous vehicles and consumer robotics.

3 EFFICIENT MODEL DESIGN

Efficient model design addresses the gap between foundational VLA complexity and resource-constrained deployment through lightweight architectures and compression techniques. Its taxonomy covers structural optimization, inference acceleration, parameter reduction, and token handling.

3 EFFICIENT MODEL DESIGN: Efficient model design combines lightweight architectures and compression techniques to reduce VLA parameters and computational costs while preserving semantic integrity.The taxonomy includes efficient architectures and model compression as complementary facets.

3.1 Efficient Architectures

Efficient VLA architectures reduce computation through attention and backbone alternatives, faster action decoding, lightweight components, sparse experts, and hierarchical execution.

3.1.1 Efficient Attention: Quadratic attention costs motivate linear-time architectures, efficient masking, and KV-cache optimization for scalable VLA inference.SARA-RT converts Transformers to linear attention, while Long-VLA uses phase-aware masking for extended operations.
3.1.3 Efficient Action Decoding: Alternative decoders replace token-by-token autoregression with parallel or generative action synthesis to reduce latency in embodied control.Generative approaches include diffusion-based continuous action decoding and four-step DDIM sampling in HybridVLA.
3.1 Efficient Architectures: Efficient architectures span attention optimization, Transformer alternatives, action-decoding acceleration, lightweight components, mixture-of-experts, and hierarchical systems.The taxonomy includes six complementary strategies for improving architectural and inference efficiency.
3.1.4 Lightweight Components: Lightweight components reduce overhead through smaller backbones and compact policy heads, including RoboMamba’s 3.7M-parameter head representing 0.1% of total parameters.TinyVLA pairs a lightweight VLM below 1.4B parameters with a diffusion policy decoder.
3.1.5 Mixture-of-Experts: Sparse expert routing activates only selected subnetworks, while hierarchical systems asynchronously separate semantic planning from rapid action execution.Hierarchical pipelines use high-level VLM representations to steer lightweight, high-cadence policies.
3.1.6 Hierarchical Systems: Hierarchical VLA pipelines combine cognitive depth with operational alacrity by decoupling compute-intensive VLM inference from latency-sensitive action generation.RoboDual uses OpenVLA for high-level planning and a lightweight Diffusion Transformer for rapid enactment.

3.2 Model Compression

Model compression improves VLA efficiency through layer pruning, quantization, and token optimization while retaining useful multimodal action capability.

3.2 Model Compression: Model compression combines layer pruning, quantization, and token optimization to reduce model depth, memory use, inference cost, and processed tokens.Token optimization includes compression, pruning, and caching.
3.2.1 Layer Pruning: Layer pruning removes redundant layers through early exits, selective skipping, or learned routing to reduce parameters and inference latency.DeeR-VLA uses action-consistency-based early exits, while MoLe-VLA dynamically bypasses superfluous layers.
3.2.2 Quantization: Aggressive 4-bit post-training quantization halves GPU memory demands while maintaining real-world robotic proficiency comparable to full-precision baselines.OpenVLA provides empirical validation for quantization in large-scale VLAs.
3.2.3 Token Optimization: Token optimization reduces quadratic inference costs by compressing, pruning, or caching visual and action tokens.SmolVLA limits visual tokens to 64 per frame, while FAST compresses action sequences using DCT and BPE.
3.2.3 Token Optimization: Token-pruning methods increasingly use task relevance, diversity, spatial semantics, instruction sensitivity, motion, or temporal continuity to select tokens.SpecPrune-VLA addresses the unreliability of relying only on local information through self-speculative pruning.
3.2.3 Token Optimization: Token caching reuses invariant key-value representations across frames or diffusion denoising steps, reducing repeated computation.VLA-Cache targets static tokens, while HybridVLA extends KV caching to diffusion models.

3.3 Discussion

Efficient VLA design combines adaptive architectures and compression to reconcile high-level reasoning with reactive execution, but efficiency can trade off against fidelity and stability.

3.3 Discussion: Linear-time attention, hierarchical decoupling, pruning, quantization, and task-aware distillation support resource-conscious VLA deployment while retaining perceptual-motor invariants.The discussion characterizes efficient design as a shift away from scaling-centric paradigms.
3.3 Discussion: Aggressive structural optimization can induce semantic drift, weaken long-horizon dexterity, destabilize asynchronous control, and limit adaptability under static importance metrics.The survey identifies hardware-algorithm co-design as a future direction for balancing efficiency and robustness.

4 EFFICIENT TRAINING

Efficient VLA training reduces resource demands through data-efficient pre-training, compact action representations, staged or reinforcement learning, and parameter-efficient adaptation.

4 EFFICIENT TRAINING: Efficient training is organized into pre-training, which creates an action-aware policy, and post-training, which specializes it for downstream tasks and environments.The taxonomy addresses computational and data overhead across both stages.
4.1.1 Data-Efficient Pre-training: Data-efficient pre-training combines scarce robotic trajectories with non-robotic data, including ego-centric videos, to reduce dependence on extensive robotic datasets.Strategies include augmenting expert trajectories, learning latent actions, and learning dynamics from action-less videos.
4.1.1 Data-Efficient Pre-training: Mixed-data co-training improves efficiency and generalization by jointly learning from heterogeneous datasets across quality and modality.Examples combine expert or suboptimal robotics data with synthetic, internet, or cross-domain information.
4.1.2 Efficient Action Representation: Compact action representations transform high-dimensional, continuous, noisy controls into latent or compressed spaces that reduce the policy-learning search space.Autoencoder-based approaches use VQ-VAE, VAE, or world-model-derived latents, while FAST reports up to a 5× reduction in pre-training time.
4.1.2 Efficient Action Representation: Cross-modal feature linkage connects human demonstrations or vision-language features to robot actions through shared action spaces and bridge attention.EgoVLA uses MANO-based action alignment, while VLA-Adapter uses a Bridge Attention module.
4.1.3 Other Pretraining Strategies: Multi-stage training, reinforcement learning, and LoRA reduce training burden by separating capabilities, improving data utilization, simulating interactions, or limiting parameter updates.LoRA preserves the backbone while enabling targeted action integration through lightweight adapters.

4.2 Efficient Post-Training

Efficient post-training adapts pretrained VLAs to downstream tasks through supervised fine-tuning and reinforcement learning while reducing data and computational demands.

Efficient Post-Training: Efficient post-training includes supervised fine-tuning for targeted parameter updates and reinforcement learning for policy optimization.
Supervised Fine-Tuning: Supervised fine-tuning specializes pretrained VLAs using labeled downstream data while preserving multimodal generalization.
Supervised Fine-Tuning: LoRA offers an effective performance-compute trade-off among five parameter-efficient fine-tuning strategies explored for OpenVLA.OpenVLA compares full fine-tuning, last-layer-only, frozen vision, sandwich fine-tuning, and LoRA.
Efficient Adaptation: Reusable skills and lightweight adaptation mechanisms further reduce data or parameter requirements for transferring VLAs to new manipulation tasks.Examples include atomic skill libraries, transfer from 50 real-world samples, and training only a learnable <ACT> token embedding.
RL-Based Method: RL-based post-training improves VLA adaptation under scarce data, with online and offline methods producing substantial success gains from few demonstrations.Reported examples include 4% to 97% success using one demonstration and 96.3% real-world success after online handover from 20–30 demonstrations.

4.3 Discussion

Efficient VLA training uses unlabeled egocentric videos, latent action modeling, and spectral compression to reduce data and policy-search costs, but alignment and adaptation limitations remain.

Discussion: Unlabeled egocentric videos provide motor priors while latent action modeling and spectral compression make policy search more compact.
Limitations: Human-video kinematic mismatch, limited adaptation flexibility, reward misspecification, and distribution shifts constrain efficient VLA training.The discussion identifies offline value overestimation and online sample-efficiency bottlenecks as additional reinforcement-learning challenges.

5 EFFICIENT DATA COLLECTION

Efficient data collection shifts VLA dataset creation from costly teleoperation toward human-supervised, simulated, curated, and computation-driven approaches, while embodiment and sim-to-real gaps remain.

Human-in-the-Loop Collection: Traditional human demonstrations are costly and difficult to scale because they depend on experts, specialized hardware, annotations, and physical deployments.The 1:1 relationship between human demonstration time and collected data further limits throughput.
Human-in-the-Loop Collection: Human involvement can be made more efficient by using language interfaces or limiting operators to supervisory intervention at failures.CLIP-RT removes the need for expert knowledge but retains continuous engagement, whereas GCENT intervenes only at failed or near-failed steps.
Simulation-Based Collection: Simulation enables automated, parallelized generation across settings, objects, lighting conditions, and embodiments at lower resource cost.Representative pipelines use photorealistic grasping data, quadruped manipulation data, analytical grasp models, and expert-simulated bimanual trajectories.
Data Reuse and Cross-Domain Sources: Existing community datasets and egocentric human videos expand available data, but require curation and methods for bridging human-robot embodiment differences.SmolVLA standardizes task descriptions and camera viewpoints, while EgoVLA treats humans as a form of robot and uses skill-rich human manipulation sequences.
Discussion: Computation-driven acquisition decouples dataset expansion from human time through simulation, autonomous exploration, and cross-domain manifold alignment.
Limitations: High-throughput synthetic and cross-modal data still face sim-to-real, embodiment, safety, and action-label limitations that constrain physical utility.The survey notes that imprecise internet-scale action labels can require continued expert supervision for high-precision tasks.

6 APPLICATIONS

Efficient VLAs support deployment across vehicles, homes, industrial systems, and medical or assistive settings by meeting constraints on latency, power, privacy, precision, and scalability.

Intelligent Vehicles and Autonomous Driving: In intelligent vehicles, efficient VLAs enable low-latency end-to-end mapping from multimodal sensory inputs to driving actions on automotive-grade hardware.They support responses to traffic-officer gestures and complex verbal navigation commands under tight computational and power budgets.
Home Robotics: In homes, edge-deployed efficient VLAs enable offline open-ended assistance, preserve data privacy, reduce power consumption, and support responsive interaction.
Industrial Robotics: In industrial settings, reduced parameters and computational overhead make advanced intelligence more economically deployable across large robot and AGV fleets.
Medical and Assistive Robotics: Medical and assistive robots require efficient architectures for real-time sensorimotor processing and local handling of sensitive health information.

7 CHALLENGES AND FUTURE WORKS

The survey identifies persistent challenges in efficient VLA design, training, and data, then proposes system-level directions for scalable, adaptive, and deployable embodied intelligence.

Challenges: Efficient VLA progress remains constrained by trade-offs among multimodal integration, scalability, robustness, and resource efficiency.The survey frames these challenges across model, training, and data dimensions.
Model: Balancing Compactness and Expressivity: Aggressive model compression can lose fine-grained spatiotemporal fidelity, while acceleration and modular routing can weaken robustness or real-time deployability.Sub-billion-parameter designs may also degrade long-horizon reasoning and adaptation to unseen tasks.
Training: Scalability versus Stability Trade-offs: Efficient training must balance computational frugality with convergence reliability amid embodiment shifts, unstable reinforcement learning, and distorted action representations.These issues hinder scalable and reproducible training across heterogeneous robotic platforms.
Data: Quality, Diversity, and Accessibility Barriers: VLA data remains limited by costly human collection, imperfect synthetic realism, augmentation bias, noisy exploration, and a lack of standardized cross-domain repositories.These barriers affect data quality, diversity, reproducibility, and access.
Model: Balancing Compactness and Expressivity: Future model designs should dynamically adapt computation through token pruning, context-aware routing, modality-agnostic backbones, and hardware-software co-design.These directions aim to preserve critical spatiotemporal information while reducing computation across task and hardware contexts.
Training and Data: Future training and data systems should become continual, decentralized, physics-informed, and generative to reduce adaptation costs and reliance on high-cost real data.Proposed mechanisms include federated learning, physics-informed objectives, meta-learning, diffusion-guided synthesis, and multi-agent exploration.

8 CONCLUSION

The survey consolidates fragmented Efficient VLA research around the model-training-data loop and organizes it into three interdependent pillars. Its synthesis highlights trade-offs between compactness and multimodal expressivity while motivating adaptive, co-designed systems.

Conclusion: The survey presents a unified taxonomy of Efficient VLAs spanning Efficient Model Design, Efficient Training, and Efficient Data Collection.The taxonomy covers the model-training-data loop and addresses computational and data bottlenecks.
Conclusion: Its synthesis exposes intrinsic trade-offs between model compactness and multimodal expressivity.The framework is intended to help navigate systemic tensions in efficient VLA design.
Conclusion: The survey’s challenges and future directions call for adaptive, co-designed systems beyond isolated efficiency optimizations.The stated roadmap targets a transition from resource-bound prototypes toward more ubiquitous physical-world intelligence.

9 BIOGRAPHY SECTION

The biography section profiles the survey’s authors, including researchers and students affiliated primarily with Tongji University, the University of Electronic Science and Technology of China, Southwest Jiaotong University, and the University of Trento.

Author Biographies: Zhaoshu Yu and Bo Wang are Tongji University students whose research interests include vision-language-action models, embodied intelligence, and multimodal learning.Yu is pursuing a B.S. degree and plans to begin a direct Ph.D.; Wang is a junior undergraduate.
Author Biographies: Pengpeng Zeng and Haonan Zhang work on visual or multimodal understanding, machine learning, and related areas including reinforcement learning.Zeng is a Tongji University researcher, while Zhang is pursuing a Ph.D. at the University of Electronic Science and Technology of China.
Author Biographies: Ji Zhang is an assistant professor at Southwest Jiaotong University whose research includes few-shot learning, transfer learning, and robotics.He earned his Ph.D. from the University of Electronic Science and Technology of China and has published in major venues.
Author Biographies: Zheng Wang is affiliated with Tongji University and focuses on vision-language action, multimedia understanding, and computer vision.He holds B.E. and Ph.D. degrees from Zhejiang University.
Author Biographies: Lianli Gao, Jingkuan Song, Nicu Sebe, and Heng Tao Shen are senior academics with research spanning multimedia, computer vision, machine learning, artificial intelligence, and big data.Their affiliations include the University of Electronic Science and Technology of China, Tongji University, and the University of Trento.