A Survey on Vision-Language-Action Models: An Action Tokenization Perspective

Yifan Zhong; Fengshuo Bai; Shaofei Cai; Xuchuan Huang; Zhang Chen; Xiaowei Zhang; Yuanfei Wang; Shaoyang Guo; Tianrui Guan; Ka Nam Lui; Zhiquan Qi; Yitao Liang; Yuanpei Chen; Yaodong Yang

Source-linked AI summary

A Survey on Vision-Language-Action Models: An Action Tokenization Perspective

Yifan Zhong, Fengshuo Bai, Shaofei Cai, Xuchuan Huang, Zhang Chen, Xiaowei Zhang, Yuanfei Wang, Shaoyang Guo, Tianrui Guan, Ka Nam Lui, Zhiquan Qi, Yitao Liang, Yuanpei Chen, Yaodong Yang

arXiv:2507.01925v1Jul 2, 2025cs.RO

TL;DR

VLA research lacks a comprehensive understanding of how action tokens are formulated and used. This survey unifies existing models through action-token categories, synthesizes their trade-offs and future directions, and concludes that raw-action end-to-end models are likely to benefit as models and datasets scale, despite practical limitations in reasoning, grounding, and embodied resources.

Problem
Existing VLA models are diverse, while action tokenization lacks a comprehensive and systematic understanding needed to interpret their designs.
Method
The survey organizes VLA research through a unified framework of modules and action tokens, reviewing token categories, representative models, strengths, limitations, data sources, and future directions.
Results
The survey concludes that raw actions are the most direct executable representation and that raw-action end-to-end VLA models are likely to evolve as models grow stronger and datasets expand.
Takeaways & Limitations
Future VLA development should combine complementary action tokens and progress toward hierarchical, agent-like systems with broader cognitive functions.
Takeaways & Limitations
Embodied deployment remains constrained by reasoning latency, brittle symbolic grounding, and the limited availability of high-quality robot data.

Abstract

from arXiv · show

The remarkable advancements of vision and language foundation models in multimodal understanding, reasoning, and generation has sparked growing efforts to extend such intelligence to the physical world, fueling the flourishing of vision-language-action (VLA) models. Despite seemingly diverse approaches, we observe that current VLA models can be unified under a single framework: vision and language inputs are processed by a series of VLA modules, producing a chain of \textit{action tokens} that progressively encode more grounded and actionable information, ultimately generating executable actions. We further determine that the primary design choice distinguishing VLA models lies in how action tokens are formulated, which can be categorized into language description, code, affordance, trajectory, goal state, latent representation, raw action, and reasoning. However, there remains a lack of comprehensive understanding regarding action tokens, significantly impeding effective VLA development and obscuring future directions. Therefore, this survey aims to categorize and interpret existing VLA research through the lens of action tokenization, distill the strengths and limitations of each token type, and identify areas for improvement. Through this systematic review and analysis, we offer a synthesized outlook on the broader evolution of VLA models, highlight underexplored yet promising directions, and contribute guidance for future research, hoping to bring the field closer to general-purpose intelligence.

Executive Summary

The survey unifies VLA models through progressively grounded action tokens and argues that future systems should strategically combine token types within hierarchical architectures. It also identifies data, hardware, reinforcement-learning efficiency, safety, and alignment as major development concerns.

VLA Unified Framework and Action Token Taxonomy: VLA models process vision and language through modules that produce progressively grounded action tokens before executable actions.The taxonomy includes language description, code, affordance, trajectory, goal state, latent representation, raw action, and reasoning.
Action Token Trends: Future VLA systems are expected to synthesize complementary token types rather than rely on one dominant representation.The survey highlights language plans, code, affordances, trajectories, goal states, latent representations, raw actions, and reasoning as complementary components.
VLA Architecture Trends: Hierarchical architectures may use language and code for long-horizon planning, then combine goal-state prediction, trajectories, and affordances before mapping to raw actions.The proposed near-term hierarchy places planning and logic control above intermediate motion representations and executable controls.
From Imitation to Reinforcement Learning: Reinforcement learning could provide trial-and-error exploration beyond imitation learning, but real-world deployment requires more efficient algorithms.High reset costs and low interaction efficiency remain practical obstacles, while VLM-generated dense rewards may accelerate training.
From VLA Models to VLA Agents: VLA research should evolve toward proactive agents with integrated memory, exploration, planning, and reflection.This shift would also replace predominantly linear processing with more complex bidirectional and graph-structured topologies.
The Triad of Progress: Model, Data, and Hardware: Embodied progress remains constrained by limited robotic platforms and scarce high-quality data, while safety and human alignment receive insufficient emphasis.The survey argues that models, data, and hardware must co-evolve for robust general-purpose intelligence.

1. Introduction

Foundation models provide broad multimodal capabilities, but their intelligence remains confined to digital environments. VLA models extend these capabilities toward physical task execution, and the survey organizes their diversity through a unified action-token framework.

1. Introduction: Foundation models learn broad, transferable capabilities from internet-scale data and generalize across complex tasks and novel scenarios.The context describes language, vision, and vision-language foundation models as versatile building blocks for multimodal systems.
1. Introduction: VLA models address the boundary between digital intelligence and physical tasks by generating actions conditioned on visual and linguistic inputs.They build upon at least one large-scale vision or language foundation model.
1. Introduction: Existing VLA systems express action guidance through plans, executable code, affordances, and scene-level trajectories.Examples include SayCan, PaLM-E, Code as Policies, object affordances, and predicted trajectories for downstream control.
1. Introduction: The survey abstracts diverse VLA architectures as sequences of modules that transform vision-language inputs into increasingly actionable tokens and executable actions.It defines VLA modules as differentiable subnetworks or functional units supporting processing and execution.
1. Introduction: The survey targets the lack of systematic understanding of action tokenization by reviewing VLA evolution, categories, design choices, and future directions.Its scope includes foundation-model background, embodied AI, and structured analysis of the VLA research landscape.

2. The Evolution of Language and Vision Foundation Models

The section traces how language, vision, and multimodal foundation models developed through scalable architectures, data, training, and deployment advances. These capabilities provide the technical foundation for embodied VLA systems, while visual models increasingly support specialized perception and generation.

2. The Evolution of Language and Vision Foundation Models: Embodied AI builds on the progression from language and vision foundation models toward integrated vision-language systems.The section frames this progression as preparation for the greater complexity of embodied VLA.
2.1. Language Foundation Models: Transformers, self-supervised pretraining, autoregressive modeling, scaling laws, and instruction tuning drove language-model capability and usability.The section also covers open-source models, parameter-efficient fine-tuning, mixture-of-experts, state-space models, test-time computation, and infrastructure optimization.
2.1. Language Foundation Models: Recent language-model progress includes stronger commercial systems, open-source alternatives, adaptive inference, and distributed training and inference techniques.These developments improve model capacity, specialization, reasoning, and deployment efficiency.
2.3. Vision-Language Models: VLA proliferation is presented as supported by progress in foundation models and data.Figure 3 uses an evolution timeline to relate these developments.
2.2. Vision Foundation Models: Vision Transformers treat images as visual-token sequences, enabling visual inputs to be processed similarly to textual inputs.This representational shift supports joint or aligned multimodal processing.
2.2. Vision Foundation Models: Specialized vision foundation models now support depth estimation, segmentation, tracking, grounding, and high-fidelity image and video generation.Examples include Depth Anything, SAM and SAM 2, SAMURAI, CoTracker, GLIP, Grounding DINO, and diffusion-based generators.

3. Overview of Action Tokens

The survey defines action tokens as semantically meaningful outputs or intermediate representations produced while VLA modules transform perception and language into executable behavior. It identifies eight principal token types and uses this taxonomy to compare VLA designs and their trade-offs.

3. Overview of Action Tokens: VLA models map raw perception to action by comprehending scenes and instructions, grounding them, planning subtasks, anticipating movement, and producing executable commands.These stages motivate the emergence of VLA modules and action tokens.
3. Overview of Action Tokens: A VLA module is a maximal differentiable subnetwork or nondifferentiable functional unit, while its outputs and meaningful intermediate generations are action tokens.The term parallels language tokens in large language models while emphasizing action-related information.
3. Overview of Action Tokens: Representative systems illustrate chains in which one module predicts a language subtask plan and another generates low-level robot commands.Hi Robot and a customized π0 are used to exemplify modules and intermediate action tokens.
3. Overview of Action Tokens: Most surveyed VLA models fit a unified framework where vision-language inputs pass through module sequences that progressively produce more grounded action tokens.This abstraction enables comparison across otherwise diverse architectures.
3. Overview of Action Tokens: The taxonomy contains eight token types: language description, code, affordance, trajectory, goal state, latent representation, raw action, and reasoning.Each type encodes task-relevant guidance differently for a given instruction and observation.
3. Overview of Action Tokens: The categories range from language descriptions and executable code to spatial affordances, temporal trajectories, predicted goal observations, latent vectors, raw controls, and reasoning expressions.These definitions distinguish the semantic, spatial, temporal, latent, executable, and decision-making roles of the representations.
3. Overview of Action Tokens: The survey reviews each category’s motivation, literature, advantages, limitations, and future directions, while Table 1 summarizes cross-category comparisons.Its tables also track preceding and succeeding modules and the transformations surrounding each token.

4. Language Description as Action Tokens

Language descriptions serve as interpretable action tokens for planning and grounding, especially in complex, long-horizon tasks. Their flexibility and foundation-model compatibility come with limitations in fine-grained control and inference latency, motivating hybrid tokenization for execution.

Progress and Key Papers: VLA research uses language descriptions to decompose instructions into semantically meaningful subtasks and connect abstract plans to physical actions.Early planners lacked perceptual grounding; later systems incorporated visual, affordance, state, and 3D information into planning.
Progress and Key Papers: External grounding modules improve plan feasibility and feedback, but remain inflexible, difficult to jointly reason with, and inadequate for fine-grained complex tasks.These limitations motivated VLM-based grounding and more integrated planning architectures.
Progress and Key Papers: Fine-grained language motions provide interpretable spatial and temporal guidance between vision-language inputs and low-level actions.They support cross-task data sharing, composition, generalization, data efficiency, and human correction across manipulation and navigation.
Advantages of Language Descriptions: Language descriptions integrate naturally with LLMs and VLMs, enabling zero-shot planning, benefiting from co-training data, and supporting complex temporally extended tasks.Co-training can transfer world knowledge into VLA models and improve generalization.
Advantages of Language Descriptions: Their interpretability supports human oversight, intervention, safety, transparency, controllability, and iterative improvement from online correction data.Language-based plans can incorporate human-in-the-loop corrections and dynamic feedback.
Discussion and Future Directions: Language descriptions are ambiguous and insufficiently expressive for fine-grained control, while large-model generation can introduce latency in dynamic or real-time settings.The survey suggests using language mainly for high-level planning and alternative tokens such as affordances, trajectories, or goal states for low-level execution.

5. Code as Action Tokens

Code-based action tokens represent executable programs that connect high-level instructions with perception and low-level robot control. They support structured planning and interpretability, but remain constrained by predefined APIs and brittle under unanticipated physical conditions.

Motivation: Code-based action tokens use executable snippets or pseudocode to represent actions for complex, long-horizon manipulation tasks.They address the limited expressiveness of discrete signals and direct language commands.
Advantages: Code provides logical structure, third-party library access, and a transparent bridge from instructions to low-level robot primitives.LLMs can synthesize task-relevant code from natural-language and visual inputs.
Representative approaches: Foundational systems map language instructions to Python code that processes perception, parameterizes robot APIs, and executes tasks.Code as Policies also integrates libraries such as NumPy for spatial reasoning.
Representative approaches: Code-based systems have been extended with prompting strategies, commonsense reasoning, grounded API naming, goal-state planning, and hybrid search.Text2Motion uses generated goal states to define task success and termination criteria.
Limitations: API dependence limits expressiveness and adaptability when dynamic or novel environments require behaviors absent from the predefined library.Unrepresented properties such as slippery surfaces or fragile objects can prevent nuanced action generation.
Limitations: Rigid symbolic representations create execution brittleness when real-world states violate API preconditions, producing safety risks despite syntactically correct code.Failures can include unsuccessful grasps, object damage, or hardware damage.
Future directions: Comprehensive multimodal API libraries could let VLMs orchestrate perception, reasoning, and robust action primitives for long-horizon tasks.The proposed libraries include object detection, tracking, spatial analysis, and reusable action functions.
Future directions: Formal verification, model checking, runtime monitoring, and human-in-the-loop refinement are proposed to improve safety, robustness, transparency, and controllability.Runtime monitoring can trigger safe shutdowns or recovery when anomalies occur, while interpretable code supports debugging and collaborative refinement.

6. Affordance as Action Tokens

Affordance tokens provide spatially grounded guidance for selecting actionable regions and assessing physical feasibility, using representations such as keypoints, boxes, masks, and maps. Their precision and efficiency trade-offs are accompanied by limitations in 3D understanding, temporal dynamics, and robustness to visual perturbations.

Overview: Affordances bridge visual perception and physical interaction by identifying actionable regions and evaluating feasibility from multimodal inputs.They abstract away embodiment-specific control mechanisms and support cross-platform interaction.
Representations: Keypoints, bounding boxes, segmentation masks, and affordance maps offer different trade-offs among interaction precision, computational complexity, and task demands.Keypoints target precise contacts, whereas boxes provide coarser localization.
Keypoints: Keypoints compactly encode a contact position and interaction direction, linking language-conditioned visual grounding to low-level control.They have supported semantic-part grounding, conditioned skills, affordance memory, and zero-shot manipulation.
Keypoints: Temporal keypoint sequences extend affordances from where to act toward how actions unfold over time in object-centric manipulation.Magma and VidBot predict instruction- and observation-conditioned keypoint trajectories.
Bounding boxes: Bounding boxes efficiently localize objects in 2D or encode their physical extent through eight 3D corner points, though they lack fine-grained geometry.They can support temporally consistent grounding and interaction inference from hand-object proximity.
Segmentation masks: Segmentation masks encode pixel-level contours and part geometry, enabling precise grounding of functional regions such as graspable or wipeable areas.Foundation models including SAM and Florence-2 improve language-conditioned segmentation quality and generalization.
Limitations: Affordance tokens often lack native 3D geometry, temporal affordance changes, and robustness to occlusion or motion blur.These weaknesses impair precise, contact-rich manipulation in dynamic, delicate, or visually challenging scenes.
Discussion and future directions: Future directions include native 3D affordances, temporal affordance dynamics, and uncertainty-aware policies robust to visual perturbations.Suggested representations include neural radiance fields, 3D Gaussian splatting, explicit meshes, and probabilistic affordances.

7. Trajectory as Action Tokens

Trajectory action tokens represent motion explicitly at increasing levels of abstraction, from sparse keypoint paths to dense scene-wide optical flow. They improve data scalability and cross-task generalization, but face limitations in 3D understanding, computational efficiency, and task suitability.

7.2. Progress and Key Papers: Trajectory tokens enable learning from off-domain videos because trajectories can be extracted without explicit action annotations.They are more explicit, explainable, and human-understandable than latent representations, facilitating training and debugging.
7.1. Overview of Trajectories: Trajectory tokens comprise Point Trajectory, Visual Trajectory, and Optical Flow, differing in abstraction and information density.Point trajectories encode keypoints, visual trajectories render paths in images or videos, and optical flow represents every pixel’s motion.
7.1. Overview of Trajectories: Point trajectories encode K critical-point paths over T steps, providing targeted and numerically precise motion guidance.They are used for vehicle waypoints and 2D paths for end-effectors or objects.
7.2. Progress and Key Papers: Trajectory-based methods support generalization across tasks and visual or semantic variations by sharing motion patterns in trajectory space.RT-Trajectory outperforms RT-1, RT-2, and RT-1-Goal on unseen tasks, while other approaches use hierarchical 2D and 3D representations.
7.4. Discussion and Future Directions: Trajectory representations often lack explicit 3D information and orientation, limiting them for non-planar and dexterous manipulation tasks.Depth supplementation mitigates some 3D ambiguity, but point trajectories still primarily encode position.
7.4. Discussion and Future Directions: Trajectory generation can be computationally expensive, while VLM-produced waypoints may be too sparse for smooth control.Traditional planners and temporal alignment modules are proposed to refine sparse predictions or execute full trajectories in real time.
7.4. Discussion and Future Directions: Trajectories work best for precise motion-path tasks but are less suitable for partial observability, complex interaction logic, force application, and affordance understanding.Their limitations reflect a trade-off between precise motion guidance and semantic richness.

8. Goal State as Action Tokens

Goal states use predicted visual outcomes as intermediate action tokens between language instructions and executable actions. They improve data scalability, task specificity, interpretability, and evaluation, while generation quality remains a central limitation.

8. Goal State as Action Tokens: Goal-state methods predict a visual representation of the intended outcome to bridge high-level instructions and low-level actions.They commonly use a high-level generative model to synthesize the goal state and a lower-level policy to produce actions.
8.1. Goal-State Forms: Single-frame goal states depict desired scenes as RGB, RGB-D, or 3D point-cloud observations.Hindsight relabeling can treat final frames from short robot-play windows as automatically generated goal annotations.
8.2. Multi-Frame Goal States: Multi-frame goal states provide temporal context and finer-grained motion information, reducing execution ambiguity.Short goal videos capture how scenes evolve and therefore provide additional how-to-do cues.
8.3. Advantages of Goal State: Goal states scale training data through hindsight relabeling and enable use of action-free videos and human-executed goals.These sources can improve real-world dynamics learning and cross-embodiment knowledge transfer.
8.3. Advantages of Goal State: Goal states provide precise spatial and visual guidance, strong interpretability, and straightforward quality evaluation against language instructions.Their explicit visual form supports human understanding, debugging, intervention, and language-image-based assessment.
8.4. Limitations and Future Directions: Generated goal states may be overspecified or inaccurate, causing lower-level policies to focus on trivial details or become overconstrained.Unnecessary precision can make task completion harder when the exact generated details are not critical.
8.4. Limitations and Future Directions: Goal states remain a promising direction as improving image and video generation provides increasingly consistent visual content for goal specification.The passage frames this as a development outlook rather than an established result.

9. Latent Representation as Action Tokens

Latent representations compress action-relevant information into learned action tokens, supporting scalable pretraining, efficient training, and expressive multimodal representations. Their central drawback is reduced explainability and controllability, making construction quality and granularity important research concerns.

9. Latent Representation as Action Tokens: Latent-action methods address scarce embodiment-specific action labels by leveraging web videos and heterogeneous cross-embodiment datasets.These sources are abundant but often lack explicit actions or exhibit embodiment gaps.
9. Latent Representation as Action Tokens: A typical latent pipeline constructs an unsupervised latent action space, trains a VLM to predict latent actions, then fine-tunes action translation for a target embodiment.The approaches are categorized as vision-based, action-based, or goal-based according to what the latent actions represent.
9.1. Vision-Based Latent Representation: Vision-based latent construction distills visual transformations into codebook vectors, whereas action-based methods encode fixed-length action chunks into skill vocabularies.Goal-based methods instead encode entire task trajectories into latent vectors representing overall goals.
9.1. Vision-Based Latent Representation: UniVLA’s latent space is 6.4% more effective than Genie’s in ablation results after using semantic features and instruction-guided disentanglement.These techniques target task-relevant tokens while reducing task-irrelevant visual variation.
9.3. Goal-Based Latent Representation: Goal-based latent representations support reasoning and long-horizon task execution in virtual open-world environments such as Minecraft.OmniJARVIS jointly models discrete latent goals with observation, instruction, memory, and thought tokens.
9.4. Advantages of Latent Representation: Latent representations improve scalability, training efficiency, and expressive power; UniVLA achieves performance comparable to OpenVLA using 4.45% of the training time.They can encode compact kinematic semantics and incorporate modalities such as tactile feedback and audio.
9.5. Limitations and Future Directions: Latent representations are difficult for humans to interpret, control, intervene in, or debug, limiting their suitability where strict safety or reliability guarantees are required.Future construction must balance fine-grained dexterous-task representation against excessive complexity and memorization.

10. Raw Action as Action Tokens

Raw actions provide the most direct executable representation for VLA models, enabling end-to-end learning with minimal action annotation. The section reviews increasingly scaled generalist architectures, datasets, and open-source systems that predict raw actions across tasks and embodiments.

Motivation: Raw actions are direct, executable representations that require limited prior human knowledge, structural constraint, and action-token annotation.Their format aligns naturally with real-world robot data and favors end-to-end learning.
Early Architectures: Early VLA systems fused vision and language features before mapping them to raw action sequences with relatively simple decoders.LangLfP combined MLP and CNN encoders with a CVAE decoder and paired goal-image-conditioned data with human language commands.
Transformer-Based Generalists: Transformer-based generalists scale datasets, task domains, and model capacity to unify multimodal inputs and robot actions.Gato trained a 1.2B-parameter decoder-only transformer on 596 control tasks totaling 1.5T tokens plus eight vision-language datasets.
Autoregressive Robot VLA: RT-2 discretizes raw robot actions into action bins, enabling autoregressive inference while retaining a foundation VLM backbone.The approach supports co-training on vision-language and robot-action data, with reported gains from cross-embodiment training on OXE’s 1M-plus trajectories across 22 robots.
Open-Source VLA: Open-source systems broaden reproducibility and adaptation, with OpenVLA providing a 970K-trajectory heterogeneous-robot dataset, codebase, and lightweight robot-specific fine-tuning.OpenVLA reports efficient adaptation using 10 to 150 trajectories per robot and also explores parameter-efficient fine-tuning and quantized inference.

10.4. Video Pretraining and Robot Data Fine-Tuning

Video pretraining and diffusion-based action generation extend raw-action VLAs by modeling temporal dynamics and producing action chunks. These approaches improve generalization and control frequency, but introduce latency, boundary consistency, and representation-alignment challenges.

Video Pretraining: Video generative pretraining learns future visual dynamics before robot-data fine-tuning, supporting action generation in simulation and real-robot settings.GR-1 uses a GPT-style transformer for future-frame prediction and then fine-tunes on robot datasets.
Diffusion-Based Action Chunking: Diffusion and flow-based action heads address limitations of discrete autoregressive tokenization for continuous or multimodal actions.Octo uses a diffusion head, while π0 combines flow matching with action chunking.
Diffusion-Based Action Chunking: π0 combines broad pretraining with task-specific fine-tuning to achieve strong zero-shot generalization and high performance on complex multistage tasks.Reported tasks include laundry folding, box building, and egg packing, with minimal fine-tuning data.
Diffusion-Based Action Chunking: 50 Hz is the reported control frequency for π0, an order-of-magnitude improvement over RT-2’s 5 Hz.The comparison is explicitly stated for control frequency, not overall task performance.
Limitations: Diffusion action chunking still faces inference latency and outdated-observation problems, while multiple modes can create discontinuities at chunk boundaries.These issues arise when the next chunk is generated while the robot executes the previous one.
Raw-Action Trade-offs: Raw-action approaches remain costly to scale because physical data collection is expensive, embodiment-specific, and vulnerable to catastrophic forgetting during adaptation.The survey contrasts these constraints with the greater accessibility of language data.

11. Reasoning as Action Tokens

Reasoning tokens externalize deliberative thought between multimodal instructions and executable actions, targeting complex, long-horizon embodied tasks. The survey finds benefits for instruction grounding and generalization, alongside speed and data limitations that favor lower-frequency applications.

Definition: In VLA models, reasoning is an explicitly externalized thinking process that helps generate subsequent executable action tokens.Unlike movement-oriented tokens, reasoning tokens serve an intermediary role.
Origins: Embodied reasoning builds on chain-of-thought methods and can combine instructions, scene descriptions, and action feedback for recursive language planning.Inner Monologue exemplifies this setup with an LLM and additional visual-scene and feedback modules.
VLM Integration: VLM-based reasoning reduces the need for numerous additional modules by leveraging pretrained linguistic and visual knowledge.ECoT and RAD tailor VLMs to embodied reasoning through fine-tuning or retraining.
Data Construction: Automated synthesis pipelines expand reasoning-data collection beyond costly human annotation by structuring reasoning from robot trajectories and human videos.ECoT synthesizes fixed reasoning sequences, while RAD also uses hand tracking from action-free human videos.
Advantages: Reasoning improves the instruction-action bridge, interpretability, and generalization to unseen objects and scenes in complex manipulation tasks.ECoT is reported to improve tasks requiring object identification, enumeration, and commonsense selection.
Limitations: Reasoning increases inference time, can reduce execution speed, and depends on costly high-quality datasets and manually fixed reasoning steps.ECoT’s asynchronous execution speeds inference by around 40%, but the survey calls for further acceleration.
Scope and Future Directions: Reasoning is best suited to complicated, long-horizon deductive tasks with multiple subtasks and relatively low-frequency execution demands.This suitability follows the survey’s stated inference-speed limitation.

12. Scalable Data Sources

Scalable VLA training depends on combining web and human video, synthetic and simulated experience, and real robot data in a data-pyramid structure. Each source contributes differently, while real-world collection and sim-to-real transfer remain central bottlenecks.

Data Pyramid: VLA action-token learning requires jointly grounded visual-linguistic data, fine-grained action supervision, and embodiment-aligned sensorimotor control.The data pyramid organizes these sources by decreasing quantity and increasing embodiment specificity.
Bottom Layer: Web and Human Video: Web and human video provide visual-linguistic grounding, world modeling, temporal prediction, and coarse affordance or latent-dynamics priors.Egocentric viewpoints can reduce the embodiment gap by approximating robot perspectives.
Middle Layer: Synthetic and Simulation Data: Simulation and synthetic data offer scalable, structured, task-centric experience for learning compositional skills and robust control policies.Procedural systems such as MimicGen, DexMimicGen, and RoboCasa augment or synthesize datasets from limited expert demonstrations.
Middle Layer: Synthetic and Simulation Data: Interactive simulators support large-scale reinforcement and imitation learning, especially for high-risk scenarios requiring recovery capabilities.Their usefulness is constrained by discrepancies in visual fidelity and physics modeling relative to the real world.
Top Layer: Real Robot Data: Real robot data supplies direct supervision for physically grounded policies by capturing dynamics, sensory noise, and unpredictable environmental variation.It is described as indispensable for bridging the sim-to-real gap and learning embodiment-specific constraints.
Cross-Embodiment Data: OXE aggregates over 1 million manipulation episodes across 22 robots to support cross-embodiment policy transfer.Multi-embodiment datasets complement task-specific datasets that target fine-grained and long-horizon skills.
Data Limitations: Real-world data collection remains a bottleneck because it is costly, operationally complex, and slow, motivating pretraining on simulation or web-scraped data.The survey presents this mixed-data strategy as prevalent for learning generalizable representations before real-robot adaptation.

13. General Discussions and Future Directions

The survey finds no dominant action-token paradigm and argues that complementary token types should be strategically combined in hierarchical VLA architectures. It also identifies reasoning, agentic organization, reinforcement learning, safety, and robot-data limitations as central future directions.

Trends of Action Tokens and VLA Models: No single action-token type dominates, so VLA development should combine complementary tokens across hierarchical levels.Language plans and code support long-horizon planning and logical control; affordances, trajectories, and goal-state prediction support intermediate motion representations.
Trends of Action Tokens and VLA Models: Latent representations remain excluded from the proposed architecture because granularity, semantic comprehensiveness, and task-centric alignment pose training challenges.The survey therefore favors explicit action tokens that are easier to train and inspect and offer greater interpretability.
Trends of Action Tokens and VLA Models: Reasoning is proposed as an action token that should adapt computation to task complexity and operate throughout the VLA hierarchy.Current reasoning in VLA models is described as rudimentary and generally limited to relatively simple tasks.
From VLA Models to VLA Agents: Future VLA systems may evolve into agents with memory, exploration, planning, and reflection, using adaptive non-linear module and token organization.The survey contrasts this agent-centric direction with current linear processing and notes world models as a way to enhance planning and online exploration.
Training, Safety, and Data: Imitation learning is limited by demonstrator capabilities and suboptimal demonstrations, while real-world reinforcement learning faces high reset costs and low interaction efficiency.The survey also identifies physical safety risks and severe robot-data shortages across quantity, modality coverage, and dexterous-data quality, motivating simulation and improved multimodal collection systems.
Training, Safety, and Data: The survey recommends combining simulation and internet-scale resources with versatile, reliable, multimodal, in-the-wild data collection systems to address robot-data bottlenecks.These directions respond to robot data requiring manual collection and remaining far less abundant than vision-language data.

14. Conclusion

This survey reviews VLA models through action tokenization, systematically examining token categories, representative models, strengths, limitations, scalable data sources, and future research directions.

Conclusion: The survey presents a comprehensive review of VLA research organized around action tokenization.It examines representative models for each token category and analyzes their strengths, limitations, and future investigation directions.
Conclusion: The survey summarizes major efforts in scalable data sources to inform and support ongoing VLA research.
Conclusion: The survey aims to guide future research on VLA models and embodied intelligence.