Do Robots Really Need World Models? CoRL 2026 Takes On World Models vs. End-to-End VLA Head-On

This is an English translation of the original Japanese article on the author’s Hatena blog: ロボットに「世界モデル」は本当に必要なのか？──CoRL 2026で真正面から議論されるWorld Model vs End-to-End VLA.

At CoRL 2026 (Conference on Robot Learning), to be held in Austin, Texas, in November 2026, there will be a rather interesting workshop.

Its title is:

Do Robots Need World Models? A Debate on Dynamics Learning, Planning, and End-to-End Robot Learning

In other words, it is an event that tackles head-on the question:

“Do robots really need world models?”

The official site is here:

https://do-robots-need-world-models.github.io/

What I personally find interesting is that it does not simply discuss “how to build a world model,” but takes the discussion all the way back to:

Do we need to build an explicit world model in the first place?

The world-model camp vs. the end-to-end camp

The basic idea of a world model is to give the robot a model that predicts how the world will change as a result of its actions, before it acts:

Current state
    ↓
Action
    ↓
Predict how the world will change
    ↓
Choose the action that leads to the desired future

Video generation models, latent world models, dynamics models, learned simulators, and World Action Models all fall into this line of work in a broad sense.

Recent VLA (Vision-Language-Action) models, on the other hand, are moving in an end-to-end direction:

Vision + Language
        ↓
       VLA
        ↓
      Action

Here, the underlying idea is:

Even without building a separate world model, perhaps the structure of the physical world and the ability to predict it can be learned in the policy’s internal latent representations.

This is exactly one of the major points of contention at this workshop.

Physical Intelligence on the “Against World Models” side

Particularly interesting is that Chelsea Finn, co-founder of Physical Intelligence, is among the speakers.

On the official site, she is listed as a speaker on the side of:

Against world models / implicit / end-to-end

Given the direction Physical Intelligence has taken, this is very interesting.

Physical Intelligence develops generalist VLAs such as π0 and π0.5, pursuing the approach of learning policies that adapt to a wide range of robots, tasks, and environments from large amounts of diverse robot data.

Rather than an architecture like:

Build a separate world model
        ↓
Predict the future
        ↓
Planner
        ↓
Action

the idea is closer to:

Large-scale robot experience
        ↓
Generalist VLA
        ↓
Acquires physical knowledge internally
        ↓
Action

That is:

Even without explicitly providing a separate model for predicting the world, perhaps the necessary understanding of the world can be acquired inside a large-scale policy.

However, some caution is needed here.

This event is a “debate.”

So it would be a bit too strong to conclude that “Chelsea Finn or Physical Intelligence completely rejects world models themselves.”

More accurately, I think it is best seen as raising the question:

Is it necessary to have an explicit world model as a module separate from the policy?

The points of debate are quite interesting

Looking at the official site, the discussion is divided into several motions.

Among them, the following questions seem particularly important:

Topic	Question
End-to-End	Is it sufficient to give the policy an internal predictive capability?
Generalization	Do explicit world models help with generalization to unseen goals?
Scaling	Do general-purpose policies scale more easily than general-purpose world models?
Safety	Are world models necessary for safety and policy evaluation?
Architecture	Is the real question not whether a world model is needed, but where to put it?
Benchmark	Do current benchmarks correctly measure the value of world models?

I think the point that is especially important is this one:

The question is not whether to use a world model, but where to put it.

A world model does not have to sit in front of the policy

When we hear “world model,” we tend to picture classical model-based control like this:

Robot
  ↓
World Model
  ↓
Planner
  ↓
Action

But a world model does not necessarily have to be placed in that position.

For example, it can be used in many different ways:

Training Data Generator
Simulation
Policy Evaluation
Failure Prediction
Safety Monitor
Representation Learning
Planning

In other words, rather than a binary choice of

World Model vs. VLA

it seems to have become a design problem of

Where in the robot system should the world model be placed?

Relation to my Vision-Tactile JEPA research

This debate is also closely related to the research theme I am working on:

Vision-Tactile JEPA World Models for Failure-Aware Long-Horizon Robotic Manipulation

What I expect from a world model in my research is not generating realistic future images per se.

For example:

Vision
  +
Tactile
  +
Action
   ↓
JEPA World Model
   ↓
Future Latent State
   ↓
Slip / Contact Loss / Failure

What matters is:

Predicting, in latent space, states that will lead to future failure.

That is, rather than using the world model as a policy that generates actions itself, I use it as:

A failure predictor that predicts the consequences of actions.

For example, an architecture like:

VLA
 ↓
Action Proposal
 ↓
Vision-Tactile World Model
 ↓
Failure Prediction
 ↓
Execute / Replan / Recovery

In this case, the VLA and the world model do not compete.

Rather, the roles can be divided as follows:

VLA: generates actions
World model: predicts the outcomes of actions
Failure model: evaluates failure risk
Planner: replans when necessary

World models may show their value in long-horizon manipulation

For short tasks, an end-to-end policy of

Observation → Action

may well be sufficient.

In long-horizon manipulation, however, problems like this occur:

Action 1
 ↓
Action 2
 ↓
Action 3
 ↓
Small errors
 ↓
The state drifts
 ↓
Failure at Action 8

If we can predict

Current state
  ↓
Candidate action
  ↓
Future latent state
  ↓
Failure risk

then it may be possible to

execute recovery or replanning before the failure occurs.

In particular, for things like:

Slip
Contact state
Grasp state
Object pose drift
Gripper failures

tactile sensing, not just vision, may be effective.

This is also a point I want to verify with my Vision-Tactile JEPA world model.

It may stop being “World Model vs. VLA”

In future Physical AI, rather than a simple competition of

World Model
vs
VLA

systems may evolve toward an architecture like this:

            Foundation VLA
                  │
        ┌─────────┼─────────┐
        ↓         ↓         ↓
      Action    Memory   World Model
                           │
                     Failure / Safety

End-to-end VLAs have the great appeal that their capabilities can be improved by scaling data and model size.

On the other hand, when considering:

Long-horizon tasks
Unknown environments
Failure recovery
Safety
Contact-rich manipulation

there is also value in a world model that can be explicitly queried with:

“What will happen next if I execute this action?”

Summary

What makes this CoRL 2026 event interesting is that it squarely debates

not “how to build a world model,” but “whether a world model is needed at all.”

The end-to-end VLA direction, as pursued by Physical Intelligence, and the predictive-model direction, such as JEPA and World Action Models, may at first glance seem to be in opposition.

Ultimately, however, the roles may be divided like this:

Policy
+
World Model
+
Memory
+
Planner
+
Failure Detection

In that sense, the truly important question may not be the yes/no question

Do Robots Need World Models?

but rather

Where should the World Model live?

In my own research, too, I believe one important direction is to position the world model not as a giant generator of future images, but as

a model that predicts future latent states and failures from vision and touch, and assists the VLA’s actions.

Where will Physical Intelligence’s end-to-end VLAs and JEPA-style world models intersect?

This debate at CoRL 2026 is an event I want to watch closely as I follow Physical AI and VLA research.

References

Do Robots Need World Models? https://do-robots-need-world-models.github.io/
Physical Intelligence https://www.pi.website/

Original Japanese article: ロボットに「世界モデル」は本当に必要なのか？──CoRL 2026で真正面から議論されるWorld Model vs End-to-End VLA