VLA vs World Models: Which Will Prevail?

Between 2025 and 2026, two distinctly different technical directions have emerged in robot AI. One is the VLA (Vision-Language-Action) approach, represented by RT-2, OpenVLA, and pi-0. The other is the World Models approach, represented by DreamerV3, Genie, and DIAMOND.

Many colleagues have asked me: between these two directions, which one should I bet on? My answer is: the question itself is wrong.

In today’s article, I want to break down and compare these two approaches, explain the logic, advantages, and bottlenecks of each, and then discuss why I believe they will ultimately converge.

The Core Logic of Each Approach

Let’s start with VLAs.

The VLA approach is straightforward: turn the robot’s control problem into a “look and describe” problem. Give the model a camera image, along with a language instruction (e.g., “put the red cup on the table”), and have it directly output how the robot’s joints should move.

Essentially, a VLA is an end-to-end perception-to-decision mapping. It doesn’t care what the physical laws are or how to solve dynamics equations. It cares about only one thing: given what I see right now and the language instruction, what is the optimal action?

This approach draws inspiration from the success of large language models. GPT can “look at images and describe them” — so why not have a similar model “look at images and take actions”? RT-2 (Google DeepMind, 2023) first proved this was feasible by discretizing robot actions into tokens and directly predicting them with a Transformer. From then on, VLA became a hot research direction.

The world model approach is completely different.

As I’ve explained in detail in previous articles, the core of a world model is learning “if I take this action, how will the environment change.” Instead of directly mapping observations to actions, it first builds an “internal simulator” of the environment, then imagines the consequences of different actions within that simulator, and finally selects the best course of action.

To use an analogy: a VLA is like an experienced driver who sees the road and immediately turns the steering wheel — relying on intuition trained from massive amounts of driving data. A world model is like a cautious new driver who, before every lane change, mentally simulates: “If I change lanes, how will the cars behind me react? Is it safe?” — relying on understanding of the physical world.

One-sentence summary of the difference: VLAs “see and act,” world models “think then act.”

VLA Strengths and Bottlenecks

The biggest advantage of VLAs is the data flywheel.

The internet contains vast amounts of vision-language data (images, videos, text), all of which can be used to pre-train VLA models. Then, a small amount of robot manipulation data is used for fine-tuning. This gives VLAs inherently strong generalization — they can understand unseen objects, follow natural language instructions, and even perform simple reasoning.

Several representative works that emerged in 2025 pushed these advantages to the extreme:

pi-0 (Physical Intelligence, 2024): Uses flow matching to generate continuous action sequences, excelling at complex household tasks like folding clothes and loading dishwashers. It demonstrated that pre-trained VLAs can quickly adapt to new tasks with minimal task-specific fine-tuning.

OpenVLA (Stanford/UC Berkeley, 2024): An open-source VLA model fine-tuned from the Llama-2 vision-language model, achieving strong performance on the Open X-Embodiment dataset. Open-sourcing enables small and medium teams to participate in VLA research.

RT-2 (Google DeepMind, 2023): The first work to turn robot actions into tokens, demonstrating that a 55B-parameter VLA can zero-shot generalize to unseen objects and instructions.

But VLAs have a fundamental bottleneck: lack of long-horizon planning capability.

Every VLA decision follows the pattern “see current frame -> output action.” There is no long-term “plan.” For simple single-step tasks (grasping a cup), this works fine. But for complex tasks requiring multi-step reasoning (“tidy up this room” — need to clear the desk, then sweep the floor, then wipe the windows), VLAs tend to lose their way.

Another issue is sample efficiency. VLAs require massive amounts of robot manipulation data for fine-tuning. Collecting this data is expensive — each task requires real robots to repeatedly execute, or simulation-generated data with transfer. For small and medium teams, this is a very high barrier.

World Model Strengths and Bottlenecks

The biggest advantages of world models are sample efficiency and planning capability.

Because a world model learns the dynamic laws of the environment, it can generate large amounts of training data through “imagination.” A well-trained world model running a day of “imagination” can produce more data than a real robot running for a month. This dramatically reduces dependence on real data.

DreamerV3 has demonstrated this: with the same set of parameters, training from scratch across 55 Atari games, it achieves high-level performance with only a small amount of real interaction. On robotic manipulation tasks, its sample efficiency is 10–100x higher than traditional RL.

Planning capability is another trump card for world models. Because they can “preview” the future in imagination, world models are naturally suited for long-horizon planning. For example, if a robotic arm needs to complete a multi-step task like “open drawer -> take out tool -> tighten screw,” a world model can simulate the entire process in imagination first, find the optimal action sequence, and then execute it.

In 2025–2026, the world model direction has also seen several important advances:

DIAMOND (2024): Uses a diffusion model as a world model, surpassing DreamerV3’s performance on Atari. It demonstrated that the generative model paradigm can also be used for environment modeling.

Genie 2 (Google DeepMind, 2024): Learns interactive 3D environments from video, generating interactive virtual worlds from a single image. While not yet perfect, the direction has enormous potential.

But world models also have clear weaknesses.

First, scaling is difficult. World models need to learn the dynamic laws of the environment, which means they must have a very precise understanding of the physical world. This is much harder than VLA’s “pattern matching.” Most current world models are still working on lab-level simple tasks, far from large-scale application.

Second, weak language grounding. World models learn vision-action mappings; they don’t naturally understand natural language instructions. Making a world model follow an instruction like “put the red cup on the blue table” requires an additional language module, and there’s no good solution yet.

Third, long-term prediction error accumulation. The further into the future a world model predicts in imagination, the larger the error. This limits its planning depth — it can typically only reliably predict a few dozen steps ahead. For complex tasks requiring hundreds of steps, planning quality degrades significantly.

Comparison: A Table to Clarify the Differences

Let’s put the two approaches side by side:

Looking at this table, you might think VLAs are overwhelmingly superior. In the short term, that’s arguably true — VLAs are easier to engineer, easier to fund, and easier to demo. But world models’ advantages in sample efficiency and planning capability are things VLAs will struggle to match in the near term.

My Take: Not Who Replaces Whom, But Who Combines with Whom

After all this comparison, let me share my own perspective.

I believe VLAs and world models are not competitors — they are complementary. The ultimate direction will inevitably be their convergence.

Why? Because each one solves the other’s weakest link.

VLAs excel at generalization and language grounding but lack planning. World models excel at planning and sample efficiency but are weak at generalization and language understanding. If you combine them — using a world model for “imagination” and planning, and a VLA for perception and language understanding — you get a system that has both intuition and deliberate reasoning.

In fact, this trend is already emerging.

Although pi-0 is classified as a VLA, it uses flow matching to generate action sequences — which is essentially a form of “imagination.” It doesn’t simply output a single action but generates an entire action sequence and continuously corrects during execution. This already has the shadow of a world model.

Although Genie 2 is classified as a world model, its way of learning from video is very similar to how VLAs learn perception-action mappings from data. The technical boundary between the two is blurring.

My prediction: in 2026–2027, we’ll see an increasing number of “VLA + World Model” hybrid architectures emerge. Specifically, a possible architecture would be:

• Perception Layer: The VLA handles visual understanding and language grounding, converting camera feeds and language instructions into structured task representations
• Planning Layer: The world model reasons about multiple options in imagined space and selects the optimal action sequence
• Execution Layer: The VLA converts the planning layer’s abstract actions into concrete joint control signals

This three-layer “perception-planning-execution” architecture draws from human cognition (System 1 fast thinking + System 2 slow thinking) and also makes engineering sense.

Advice for Practitioners

If you’re an engineer or graduate student looking to enter this direction, my advice is: don’t bet on just one side.

Learn VLAs, but also understand world model principles. Learn world models, but also follow the latest VLA developments. The two approaches have significant overlap in their tech stacks — Transformer architectures, visual encoders, RL fundamentals — these are all transferable.

Specifically:

If you lean toward engineering, start with VLAs. OpenVLA is open-source, datasets are publicly available, and the tech stack heavily overlaps with LLMs, making it easier to get started. First get a VLA pipeline running, then gradually add world model planning capabilities.

If you lean toward research, start with world models. There’s more academic space in this direction — how to improve imagination-space accuracy, how to do more efficient imagination training, how to introduce language grounding into world models — these are all open problems with room for impactful results.

If you want to start a company, focus on the intersection of the two. For example: using world models to generate synthetic data for VLA training, or using VLA perception capabilities to enhance world models’ environment understanding. These directions are closer to productization and competition isn’t as fierce yet.

Final Thoughts

Debates over technical directions are nothing new in AI history. CNN vs RNN, GAN vs VAE, on-policy vs off-policy — looking back, what usually wins is not one side or the other, but the fusion of both.

VLAs and world models are the same. They’re not enemies — they’re two pieces of a puzzle. Whoever puts the two pieces together first will build truly powerful robot AI.

As an engineer working in this field, I feel fortunate to witness this process. Whether VLAs ultimately dominate, world models dominate, or the two converge, the core technical skills are transferable. Stay learning, stay curious, and you won’t go wrong.

In the next article, we’ll discuss the career prospects of embodied AI and reinforcement learning: is this direction worth investing in? Where do practitioners go from here? Stay tuned.