The emergence of Vision-Language-Action (VLA) models represents a pivotal shift in the pursuit of Embodied AI, aiming to create intelligent agents that can seamlessly translate multimodal perception and language understanding into physical action in the real world. These models bridge the historical divide between high-level reasoning systems and low-level robotic control, offering a unified, end-to-end framework for task execution. This review provides a systematic examination of the core technical components within VLA models: action representation and generation strategies. We trace their evolution, analyze the fundamental trade-offs between precision, diversity, and efficiency, and discuss emerging challenges and future opportunities in building more capable and reliable embodied AI robots.
1. Introduction: The Paradigm of Vision-Language-Action Models
Embodied AI, which focuses on agents that perceive, reason, and act within physical environments, is considered a critical pathway toward Artificial General Intelligence (AGI). The core challenge lies in grounding abstract knowledge and instructions in concrete, physical interactions. Traditional robotic pipelines often consist of fragmented modules for perception, planning, and control, which struggle with generalization and adaptability in unstructured settings. The recent success of large-scale models, particularly Large Language Models (LLMs) and Vision-Language Models (VLMs), has inspired a new class of architectures: Vision-Language-Action (VLA) models. These models integrate visual perception, linguistic understanding, and motor control into a single, cohesive framework, enabling language-conditioned robotic task execution with unprecedented flexibility.
The canonical architecture of a VLA model typically comprises three key components: a visual encoder, a language encoder, and an action decoder or policy. The visual encoder, often a pre-trained Vision Transformer (ViT), processes raw pixel inputs to extract features representing objects, scenes, and spatial relationships. The language encoder, based on an LLM, interprets natural language instructions. The action policy then maps the fused visual-language representation to executable motor commands. This end-to-end learning paradigm allows embodied AI robots to leverage internet-scale knowledge for real-world problem-solving.
Within this framework, two interrelated elements are paramount for successful embodiment: action representation and generation strategy. Action representation defines what the model outputs—the format and semantics of the control signal (e.g., joint angles, end-effector poses, discrete commands). The generation strategy defines how the model produces this output from its inputs—the algorithmic process for mapping perception to action. The design choices for these components critically influence the model’s precision, speed, robustness, and ability to handle the multimodality inherent in embodied AI tasks.
2. Action Representation in VLA Models
Action representation is the foundational layer that connects the model’s internal reasoning to the actuator commands of an embodied AI robot. It must contend with the high-dimensional, continuous nature of robotic action spaces while capturing the potential diversity of valid solutions for a given task. The research community has largely converged on two principal paradigms: discrete (tokenized) and continuous action representations.
2.1 Discrete Action Representation
Discrete action representation was a groundbreaking approach in early VLA development. Its core premise is to transform the continuous control problem into a sequence modeling problem, leveraging the immense power of Transformer architectures used in LLMs. This is achieved by quantizing or “binning” each dimension of the continuous action space into a finite set of discrete intervals, each assigned a unique token ID. For example, a 7-degree-of-freedom (DoF) robot arm’s joint angles might each be quantized into 256 bins, turning a 7D continuous vector into a sequence of 7 discrete tokens from a vocabulary of 256.
The probability of generating a sequence of action tokens $a_{1:L}$ can be modeled autoregressively using the chain rule:
$$p(a_{1:L} | o, l) = \prod_{t=1}^{L} p(a_t | a_{<t}, $l$="" $o$="" and="" instructions.
This paradigm enabled the direct application of pre-trained sequence models to robotics. Pioneering models like RT-1 demonstrated that a standard Transformer could learn hundreds of manipulation tasks by predicting sequences of discretized actions. Subsequent models like RT-2 showed that a large VLM, when fine-tuned on robot data (“co-fine-tuning”), could directly output action tokens, effectively transferring web knowledge to physical control.
| Year | Model | Core Paradigm | Platform | Action Space | Dim. | Bins | Key Insight/Limitation |
|---|---|---|---|---|---|---|---|
| 2022 | RT-1 | Imitation Learning | Mobile Manipulator | EE Pose + Base | 11 | 256 | Scaled Transformer control; quantization error. |
| 2022 | Gato | Generalist Agent | Multi-platform | EE Velocity | 5 | 1024 | Unified tokenization; slow autoregressive inference. |
| 2023 | RT-2 | VLM Co-fine-tuning | Mobile Manipulator | EE Pose + Base | 11 | 256 | Knowledge transfer from web; high compute cost. |
| 2024 | OpenVLA | VLM Parameter-Efficient FT | Various Arms | EE Pose | 7 | 256 | Open-source, strong performance; inference latency. |
| 2025 | Humanoid-VLA | Language-Motion Alignment | Humanoid Robot | Full-Body Poses | 24 | 1024 | Extends to high-DoF control; relies on RL for low-level stability. |
Advantages: Unifies robotics with powerful sequence models, enables knowledge transfer from VLMs, and offers stable training.
Disadvantages: Inherent quantization error limits precision for delicate tasks. Autoregressive generation is inherently sequential, creating a bottleneck for high-frequency control required by embodied AI robots.
2.2 Continuous Action Representation
To overcome the precision limits of discretization, continuous action representation operates directly in the native action space of the embodied AI robot. The central challenge here is multimodality: for a given task and observation, there can be multiple equally valid action trajectories. A simple regression model trained with Mean Squared Error (MSE) loss tends to average these modes, resulting in a blurry, ineffective “mean action.”
Continuous representations address this by modeling the full probability distribution over actions. Instead of predicting a single value, the model learns to generate or sample from a distribution that covers all plausible actions. This is often framed as learning a conditional distribution $p(a | o, l)$, where $a$ is a continuous vector.
Common approaches for modeling this distribution include Conditional Variational Autoencoders (CVAEs), diffusion models, and flow matching models.
- CVAE-based (e.g., ACT): Learn a latent variable $z$ that captures the mode of the action. The action is decoded from $z$ and the observation. The training involves a reconstruction loss and a KL-divergence regularization:
$$\mathcal{L}_{CVAE} = \mathbb{E}_{z \sim q_{\phi}(z|a,o)}[\|a – \pi_{\theta}(o, z)\|] + \beta D_{KL}(q_{\phi}(z|a,o) \| p(z))$$
where $p(z)$ is a prior (e.g., standard normal). - Diffusion-based: Frame action generation as a iterative denoising process. Starting from noise $a_K \sim \mathcal{N}(0, I)$, the model progressively denoises over $K$ steps to produce a clean action $a_0$. The core is learning a noise prediction network $\epsilon_{\theta}$:
$$a_{k-1} = \frac{1}{\sqrt{\alpha_k}} (a_k – \frac{1-\alpha_k}{\sqrt{1-\bar{\alpha}_k}} \epsilon_{\theta}(a_k, o, l, k)) + \sigma_k z,$$
trained to minimize $\|\epsilon – \epsilon_{\theta}(a_k, o, l, k)\|^2$. - Flow Matching: Learn a vector field $v_{\theta}(a_t, o, t)$ that defines a probability path from a simple noise distribution $p_1(a)$ to the data distribution $p_0(a)$. Actions are generated by solving an ODE: $da_t = v_{\theta}(a_t, o, t) dt$. Training minimizes a simple regression loss:
$$\mathcal{L}_{FM} = \mathbb{E}_{t, a_0, a_1} \| v_{\theta}(a_t, o, t) – (a_1 – a_0) \|^2.$$
| Year | Model | Core Paradigm | Representation Type | Key Contribution | Primary Challenge |
|---|---|---|---|---|---|
| 2023 | ACT | Imitation Learning | CVAE + Action Chunking | Precise, multi-modal action sequences for dexterous tasks. | Modeling complexity; hardware limits. |
| 2023 | Diffusion Policy | Imitation Learning | Conditional Diffusion | Introduced diffusion models for visuomotor policy, high-quality smooth trajectories. | High computational cost per inference (slow). |
| 2024 | Octo | Imitation Learning | Diffusion (DiT-based) | Scalable, generalist policy trained on Open X-Embodiment. | Performance dependent on demonstration data quality. |
| 2024 | $0\pi$ | VLM Fine-tuning | Flow Matching | Efficient, high-performance VLA model with an “action expert.” | Requires massive, high-quality demonstration datasets. |
| 2025 | DexVLA | Embodied Curriculum Learning | Multi-Head Diffusion | Plug-in diffusion expert for fine-grained dexterous manipulation. | Limited in complex, contact-rich scenarios. |
Advantages: High precision, ability to model complex multi-modal distributions, and generation of smooth trajectories ideal for embodied AI robot control.
Disadvantages: Increased model and training complexity. Generative models like diffusion can have slow, iterative inference, posing challenges for real-time control.
3. Action Generation Strategies in VLA Models
The generation strategy is the algorithmic engine that produces actions from the fused visual-language context. It directly determines the quality, speed, and reliability of the embodied AI robot’s behavior. The choice involves fundamental trade-offs between precision vs. speed and diversity vs. stability.
3.1 Autoregressive Generation
Autoregressive generation is the natural strategy for discrete action models. It produces action tokens one at a time, with each new token conditioned on all previously generated tokens and the input context. This is implemented using a Transformer decoder with causal masking.
For a sequence of $T$ actions $a_{1:T}$, the model factorizes the joint probability:
$$p(a_{1:T} | o, l) = \prod_{t=1}^{T} p(a_t | a_{1:t-1}, o, l).$$
This sequential dependency ensures temporal coherence but is the primary source of latency, as generation cannot be parallelized across time steps. While stable and effective for planning, its low output frequency (typically 3-5 Hz) is often insufficient for the high-rate control (e.g., 30-100+ Hz) needed by dynamic embodied AI robots.
3.2 Non-Autoregressive Generation
This class of strategies aims to generate entire action sequences in parallel or with minimal sequential steps, dramatically improving inference speed.
3.2.1 Parallel Decoding via Diffusion/Flow Models
Models like Diffusion Policy, Octo, and $0\pi$ generate a full action sequence (or “chunk”) in a fixed number of parallelizable steps. Unlike autoregressive models that generate time step $t$ before $t+1$, these models generate all time steps concurrently through an iterative refinement process. For a diffusion model predicting an action chunk $\mathbf{A} = [a_t, a_{t+1}, …, a_{t+H}]$, the denoising process operates on the entire $\mathbf{A}$ at each diffusion step. This allows the generation of long-horizon, smooth trajectories in a single, albeit computationally intensive, forward pass. The parallelism lies across the diffusion timesteps $k$, not the action timesteps $t$, but the output is still a full sequence delivered at once.
3.2.2 Latent Space Sampling (CVAE)
As used in ACT, this strategy decouples the multi-modal reasoning (choosing a latent mode $z$) from the action generation (decoding $z$ into a sequence). Once a latent variable $z$ is sampled from the encoder or prior, the decoder can generate the entire multi-step action chunk $\mathbf{A}$ in a single, fast forward pass: $\mathbf{A} = \pi_{\theta}(o, z)$. This is non-autoregressive in the action sequence dimension.
3.2.3 Efficient Compression and Distillation
A recent trend focuses on making powerful but slow generative models viable for real-time embodied AI robots. Techniques like Fast Action Sequence Tokenization (FAST) compress long action sequences into a few tokens in a learned latent space (e.g., using frequency domain representations), which the VLA model predicts. These tokens are then decoded back into a full-rate action sequence by a small, efficient network. Another approach is distilling a large, slow teacher model (e.g., a diffusion policy) into a small, fast student model (e.g., a single-step policy). The student is trained to mimic the teacher’s output distribution, preserving quality while drastically cutting latency.
3.3 Hybrid Generation Strategies
Hybrid strategies combine different generation paradigms within a single system to leverage their complementary strengths. A common architecture uses an autoregressive or LLM-based component for high-level task planning and a non-autoregressive component (like a diffusion model) for low-level, high-frequency control.
For instance, a model might first use an autoregressive planner to break down a language instruction into a sequence of symbolic sub-goals (e.g., [“pick apple”, “place in bowl”]). A separate, dedicated diffusion policy then takes each sub-goal and the current visual observation to generate the precise, continuous motion trajectories to achieve it. This separation allows the system to benefit from the LLM’s reasoning and long-horizon planning capabilities while maintaining the control quality and speed needed by the embodied AI robot’s actuators.
Formally, if $g_{1:M}$ is the plan from the high-level module, the low-level policy generates actions conditioned on the current sub-goal $g_i$: $a_t \sim \pi_{low}(o_t, g_i)$. This hybrid approach is exemplified by models like Groot N1 and HybridVLA, which aim to balance cognitive depth with reactive control.
4. Evaluation and Benchmarking
Progress in VLA models is measured through standardized benchmarks that test generalization, robustness, and lifelong learning. Key datasets include:
- LIBERO: Focuses on lifelong learning and knowledge transfer (procedural, spatial, conceptual). Performance is measured by success rate across task suites designed to test forward transfer, backward transfer, and area under the learning curve.
- Open X-Embodiment: A large-scale, cross-robot dataset used to train and evaluate generalist policies. It tests zero-shot generalization to new robots, tasks, and environments. The primary metric is task success rate.
Empirical results often show a trade-off. Discrete, autoregressive models (e.g., RT-2-X) show strong generalization from web knowledge but may have lower peak success rates on precise manipulation. Continuous, diffusion-based models (e.g., $0\pi$) often achieve higher success rates on benchmark tasks due to their precision but can be computationally heavier. The optimal choice depends on the specific requirements of the embodied AI robot application, balancing task complexity, available compute, and required control frequency.
5. Challenges and Future Directions
Despite rapid progress, significant challenges remain in deploying robust and general VLA models for embodied AI.
5.1 Integration with World Models
A major frontier is integrating VLA models with learned world models. While current VLAs are largely reactive, world models enable forward prediction and mental simulation. Combining them would allow embodied AI robots to perform “what-if” reasoning, plan long-horizon sequences, and recover from failures by re-planning in a learned dynamics model, moving closer to true predictive and adaptive intelligence.
5.2 Real-Time Control Efficiency
Making high-capacity VLA models run at the control frequencies of embodied AI robots (often >100 Hz) is an open problem. Future work lies in:
- Advanced Distillation & Compression: Creating ultra-efficient student models that preserve the performance of large generative teachers.
- Specialized Hardware/Software Co-design: Developing accelerators optimized for the specific compute patterns of diffusion or flow matching inference.
- Hierarchical Latent Planning: Having a slow “think” module that updates a latent plan asynchronously, while a fast “act” module executes the current plan at high frequency.
5.3 Embodiment-Agnostic Representation
The goal is to learn action representations that are abstracted away from specific robot kinematics. An ideal embodied AI robot policy would output goals in a shared semantic-geometric space (e.g., desired end-effector poses relative to objects, force profiles) that can be translated by a low-level, robot-specific controller. This would enable true “one-model-fits-all” generalization across diverse robotic platforms, from arms to quadrupeds to humanoids.
5.4 Safety and Reliability in Open Worlds
Ensuring safe operation in unstructured environments is paramount. Challenges include:
- Robust Perception: Improving visual grounding under occlusion, lighting changes, and adversarial conditions.
- Predictive Safety: Integrating collision prediction and constraint satisfaction directly into the action generation loop.
- Explainability & Debugging: Developing tools to understand failure modes in the end-to-end model, which is crucial for trust and deployment in safety-critical applications for embodied AI robots.
5.5 Beyond Imitation: Integrating Exploration and Reinforcement
Current VLA models are primarily trained via imitation learning (behavior cloning) on demonstration data, which limits their performance to the coverage of the dataset. The next leap requires incorporating principles from reinforcement learning (RL) to enable embodied AI robots to explore, optimize, and improve beyond the demonstrations. This could involve offline RL for training from mixed-quality data or fine-tuning with online interaction in simulation or the real world.
6. Conclusion
The field of Vision-Language-Action models is evolving rapidly, driven by advances in action representation and generation strategies. We have witnessed a progression from discrete tokenization with autoregressive generation towards sophisticated continuous representations powered by diffusion models and flow matching. Each approach embodies a different point in the trade-off space of precision, speed, diversity, and generality. The emerging trend of hybrid strategies seeks to combine the best of all worlds. The ultimate goal remains the creation of capable, efficient, and safe embodied AI robots that can operate autonomously in the complex, open-world environments shared with humans. Overcoming the challenges of real-time efficiency, safety, generalization across embodiments, and integration with world models will define the next chapter of research in this exciting field.