The Missing Link Between AI and Robotics?

Artificial intelligence has become very good at understanding information.

We can give an AI model some text and ask it a question. We can show it an image and ask what is happening. We can give it video, audio or documents and ask it to understand them.

But robotics has a much harder problem.

A robot cannot only understand the world.

It has to do something in the world.

It may need to pick up a cup, move a box, open a drawer, sort an object, use a tool or safely move around a room.

This is where Vision-Language-Action models, usually called VLA models, become very interesting.

A VLA model tries to connect three things:

what the robot sees, what the robot is told, and what the robot should do.

This sounds simple, but it could be one of the most important changes happening in modern robotics.

In July 2026, Google DeepMind introduced Gemini Robotics 2, its latest VLA model. DeepMind describes it as a model that can turn vision and language input into motor control, including control of robot arms and full humanoid bodies. It also introduced separate embodied-reasoning and on-device robotics models.

So are VLA models really the missing link between modern AI and robotics?

I think the answer is: possibly—but there is much more engineering required around them.

First, what is a Vision-Language-Action model?

The easiest way to understand a VLA model is to separate the name into three parts.

Vision means the robot can observe its environment.

This may come from cameras or other visual sensors.

Language means the robot can understand an instruction.

For example:

Pick up the red cup and place it on the table.

Action means the system turns that understanding into something the robot can physically do.

So the basic flow looks like this:

Camera / Vision ↓ What is around me? ↓ Language instruction ↓ What does the person want? ↓ Reason about the task ↓ Generate an action ↓ Move the robot

This is a major difference from an AI chatbot.

If I ask a chatbot:

Which cup is red?

it only needs to answer me.

If I tell a robot:

Pick up the red cup.

the system must understand the instruction, identify the correct cup, understand its position and then control the robot well enough to pick it up.

The answer is no longer just text.

The answer is movement.

How robotics traditionally works

Traditional robotics often separates a problem into different systems.

One part handles perception.

Another part handles localisation.

Another system performs planning.

A controller manages movement.

Safety systems check whether an action should be allowed.

This approach has worked extremely well and is still used throughout robotics.

Imagine a robot arm sorting coloured objects.

A traditional system might work like this:

Camera ↓ Computer vision ↓ Object detected ↓ Calculate position ↓ Choose movement ↓ Robot controller ↓ Move arm

Each stage has a specific job.

This is useful because engineers can understand exactly what each part is doing.

But it can also require a lot of engineering.

If the environment changes significantly, parts of the system may need to be changed or retrained.

A robot designed to move one type of object may struggle when it suddenly sees something completely different.

This is one area where foundation models could change robotics.

From VLM to VLA

We already have powerful Vision-Language Models, or VLMs.

A VLM can look at an image and understand it.

For example, we could show it a workbench and ask:

What objects can you see?

It might answer:

A screwdriver, two boxes, a bottle and a robotic gripper.

That is useful.

But the model is still mainly describing the world.

A VLA model goes one step further.

Instead of asking:

What can you see?

we can potentially say:

Pick up the screwdriver and place it next to the box.

Now the model needs to connect understanding with action.

That difference is important.

A VLM can understand.

A VLA is designed to understand and act.

Research reviews published across 2025 and 2026 describe VLAs as models that combine visual perception, natural-language understanding and robot action generation, with the goal of generalising across different tasks, objects and environments.

Why is this happening now?

The idea of intelligent robots is not new.

Researchers have worked on robot learning for decades.

What has changed is the quality of foundation models.

Modern AI models have learned from huge amounts of text, images and video. Because of this, they can build a broader understanding of objects, language and relationships in the world.

Robotics researchers now want to bring some of that ability into physical machines.

Instead of training a robot only for:

Move object A from position X to position Y.

we want robots that can understand instructions such as:

Clear the table.

That instruction contains much more uncertainty.

What needs to be moved?

Where should it go?

Which objects are fragile?

What should stay on the table?

What order should the robot follow?

Humans understand many of these things using general knowledge.

The hope is that foundation models can give robots more of this general understanding too.

A simple example

Imagine a robot working in a kitchen.

There is a plate, a glass, some food packaging and a cleaning cloth on a table.

You tell the robot:

Clean the table.

A traditional robot might need each part of that task defined in advance.

A more general intelligent robot could potentially break the instruction into steps:

Understand "clean the table" ↓ Look at the table ↓ Identify the objects ↓ Decide what should be removed ↓ Move objects safely ↓ Pick up the cloth ↓ Clean the surface

This requires more than object detection.

The robot needs some understanding of the meaning of the task.

This is where language models and robotics start to come together.

Embodied reasoning is also becoming important

VLA models are not developing alone.

Another important idea is embodied reasoning.

This means an AI system reasons while considering the physical world.

A normal language model may know that a box can be placed on a shelf.

A robot needs more information.

Is the shelf high enough?

Can the robot reach it?

Is there another object blocking the path?

Is the box too heavy?

Which arm should be used?

Should the robot move its body first?

This is physical reasoning.

Google DeepMind's Gemini Robotics ER 2, for example, is designed as a higher-level reasoning model. DeepMind says it can understand physical spaces, plan multi-step tasks and then hand motor execution to a lower-level VLA model.

That architecture is interesting because it separates two problems:

What should I do? ↓ Embodied reasoning

How should I physically do it? ↓ VLA / control

This could become an important pattern in future robotics systems.

VLA models could make robots easier to instruct

One of the biggest advantages could be the interface between humans and robots.

Programming a robot normally requires technical knowledge.

But language is something almost everyone already understands.

Instead of programming:

Move arm to coordinate X Rotate joint Close gripper Move to coordinate Y Open gripper

we might simply say:

Move the blue bottle into the box.

The AI system handles more of the translation between human intention and robot behaviour.

This could make robots easier to use in factories, warehouses, hospitals, farms and eventually homes.

It could also make robots more flexible.

If we need to change a task, we may not always need to completely rewrite the robot's program.

We could explain the new task using natural language.

That is a very powerful idea.

But physical AI is much harder than a chatbot

This is also where we need to be realistic.

A chatbot can make a mistake and generate a bad sentence.

A robot making a mistake can cause physical damage.

Imagine telling a robot:

Bring me the cup next to the laptop.

If the AI chooses the wrong object, that is already a problem.

But there are more questions.

Can the robot safely reach the cup?

Will it knock over another object?

How tightly should it hold the cup?

Is there hot liquid inside?

Is a person standing nearby?

The physical world introduces consequences.

This means reliability and safety matter much more.

The data problem is huge

Large language models have an enormous amount of text available for training.

Robotics does not have the same advantage.

Getting useful robot-action data is expensive.

Someone may need to operate the robot, demonstrate tasks, record sensors and cameras, label results and repeat the process many times.

Different robots also have different bodies.

A robotic arm behaves differently from a humanoid.

A drone behaves differently from both.

Researchers are therefore exploring simulation, synthetic data, human demonstrations, robot demonstrations and other ways of producing more training data.

Recent VLA research has identified data infrastructure as one of the main challenges for the field. A 2026 survey argues that future progress will depend heavily on better datasets, benchmarks and data-generation systems, not only on bigger models.

I think this is an important point.

The future of robotics AI may depend just as much on better data as it does on better models.

Latency is another major problem

Robots operate in real time.

They cannot always wait several seconds for a cloud AI system to decide what to do.

Imagine a robot holding an object while a person suddenly walks into its path.

The system needs to react quickly.

This is why on-device AI is becoming important.

Instead of sending every camera frame and decision to the cloud, some intelligence can run directly on the robot.

Google DeepMind's Gemini Robotics On-Device 2 is an example of this direction. It is specifically designed to run locally on robotic hardware and to work in situations where network latency or internet connectivity is a problem.

Running locally can provide advantages such as faster response and less dependence on a network.

But it creates another challenge:

robots have limited computing power, energy and cooling compared with large data centres.

So researchers need models that are not only intelligent.

They also need to be efficient.

Does this mean traditional robotics will disappear?

I don't think so.

There is sometimes a tendency in AI to believe that a new model will replace everything that came before it.

Robotics is more complicated than that.

A robot still needs hardware drivers.

It still needs sensor communication.

It still needs motor control.

It needs safety systems.

It may need localisation, mapping, networking and real-time communication.

Frameworks such as ROS still solve many important engineering problems.

So instead of:

VLA replaces traditional robotics

I think a more realistic future may look like:

Language + Vision ↓ AI reasoning ↓ VLA policy ↓ Planning / Safety ↓ ROS and robot systems ↓ Low-level controller ↓ Motors and hardware

In other words, AI may become another powerful layer inside the robotics stack.

It does not have to replace the whole stack.

The hybrid approach may be the practical answer

Personally, I find a hybrid architecture more convincing than giving one AI model complete control over everything.

Use AI where flexibility and understanding are needed.

Use deterministic software where exact behaviour and safety are required.

For example:

AI: Understand "put the tools away."

AI: Recognise the tools and decide where they belong.

Traditional robotics: Check whether the target position is reachable.

Safety system: Check whether a person is too close.

Motion controller: Generate stable joint movement.

Hardware: Execute the movement.

Each part is doing what it is good at.

This gives us the flexibility of modern AI without throwing away decades of robotics engineering.

What could VLA models be useful for?

If these systems continue improving, there are many possible areas.

In manufacturing, robots could adapt more easily when products or tasks change.

In warehouses, robots could understand more flexible instructions instead of following only fixed workflows.

In agriculture, a robot could combine visual understanding with instructions for inspecting, sorting or handling crops.

Assistive robots could understand natural human requests and help with everyday physical tasks.

Drones could combine language instructions with visual navigation.

Service robots could potentially handle environments that are less predictable than traditional factories.

For countries such as Bangladesh, I think areas like agriculture, manufacturing, logistics and assistive technology could be particularly interesting.

But practical adoption will depend on much more than AI models.

Cost, hardware availability, reliability, maintenance, training and safety will all matter.

Are humanoid robots the final goal?

VLA development is often shown using humanoid robots because humanoids are visually impressive and our world is designed around the human body.

Doors, stairs, shelves, tools and furniture are built for us.

A robot with a similar body may therefore be able to use existing environments without changing everything around it.

But I don't think every useful robot needs to look like a human.

A warehouse robot may be better with wheels.

A farming robot may need a completely different body.

An industrial robot arm may be better than a humanoid for a fixed manufacturing task.

A drone has completely different requirements.

The interesting part of VLA technology is therefore not only humanoid robotics.

It is the possibility of creating more general intelligence across many different types of machines.

DeepMind's recent robotics work is also exploring this multi-embodiment direction, including adapting VLA intelligence across different robot forms.

So, are VLA models the missing link?

Maybe "missing link" is slightly too strong.