What Walking Taught Me About Artificial Intelligence

What Walking Taught Me About Artificial Intelligence

I am currently working on a computer-vision project inspired by devices such as Meta’s smart glasses, but with a very different goal.

I want to build an assistive system for visually impaired people, or people with serious vision difficulties, that can understand their surroundings while they are walking and provide useful verbal guidance in real time.

For example:

• Is there an obstacle ahead?
• Is someone approaching?
• Is the road safe to cross?
• Where is the entrance to a building?
• Is there a staircase nearby?
• What is happening around the user right now?

At first, I thought about this mainly as an engineering problem.

Cameras. Computer vision. Object detection. Depth estimation. Tracking. Navigation. Language models. Speech.

Then, while randomly walking through my college campus one day, I noticed something that completely changed the way I was thinking about the problem.

I wasn't really thinking about walking at all.

I was simply… walking.

My brain was not constantly telling me:

Move your right leg forward.

Maintain balance.

Estimate the distance to that person.

Avoid that small object.

Check the road.

Adjust your direction.

Move your left leg.

None of that was happening consciously.

Instead, I was thinking about completely different things while my body continuously handled an enormous amount of information in the background.

And that realization was kind of mind-blowing.

Something Simple for Humans Is Not Simple at All

Walking feels like one of the simplest things we do.

But from an intelligence perspective, it is anything but simple.

While walking, the human brain is constantly processing visual information, estimating depth, identifying objects, understanding motion, predicting trajectories, maintaining balance, controlling muscles, remembering the environment, recognizing danger and deciding where attention should be focused.

And it does all of this continuously.

Imagine trying to build the same thing artificially.

A machine would need perception.

It would need spatial understanding.

It would need memory.

It would need prediction.

It would need decision-making.

If it had a robotic body, it would also need precise motor control, balance and continuous feedback from the environment.

And all of this has to happen quickly enough that the machine does not simply walk into a wall.

Suddenly, "just walking" doesn't look so simple anymore.

This becomes even more fascinating when we think about something as ordinary as moving our fingers.

We barely think about it.

Yet creating robotic hands with the adaptability, sensitivity and dexterity of a human hand remains an extremely difficult engineering problem.

Humans routinely perform tasks that would require enormous amounts of engineering to reproduce artificially.

The Human Brain Is an Incredible Learning System

This made me appreciate something else.

Human beings are not born knowing how to walk through a crowded street.

A child learns.

The brain observes.

It experiments.

It fails.

It adjusts.

It remembers.

Eventually, many of those learned behaviors become almost automatic.

If you look at it from a machine-learning perspective, the human body and brain form an extraordinary adaptive learning system.

If you believe in God, there is something almost amusing about thinking of God as the ultimate Human Intelligence Researcher or perhaps the greatest Learning Systems Engineer imaginable.

A system was created that can learn from surprisingly little experience, continuously adapt to new environments, operate with limited energy, combine information from multiple senses and eventually perform incredibly complicated tasks without consciously thinking about every individual step.

Meanwhile, we are trying to reproduce fragments of those capabilities using GPUs, neural networks, sensors, enormous datasets and years of research.

And we are still nowhere close to reproducing the complete system.

That is fascinating.

Maybe AI Research Is Partly About Understanding Ourselves

The deeper I go into AI and computer vision, the more I feel that we are not simply trying to "make machines intelligent."

In some ways, we are trying to formalize abilities that humans perform naturally.

How do we recognize something?

How do we understand space?

How do we know which information matters?

How do we remember?

How do we predict what happens next?

How do we learn from experience?

How do we understand another person's intention?

How do we interact with a constantly changing world?

Humans usually don't need mathematical explanations for these things before doing them.

We just do them.

But if we want machines to perform similar tasks, suddenly we have to break intelligence into components, define them, measure them and somehow recreate them computationally.

That is one reason AI becomes so interesting.

In trying to build intelligence, we are also forced to investigate what intelligence actually is.

And Then Comes the Slightly Uncomfortable Question

There is another side to this thought.

Humans became the dominant intelligent species on Earth.

Our intelligence allowed us to build civilizations, medicine, science, transportation and technologies that would have been unimaginable to previous generations.

It has produced extraordinary things.

But the same intelligence also gave us the ability to reshape the planet at enormous scale.

We consume resources, transform ecosystems, build industries and sometimes optimize for our own objectives without fully considering the long-term consequences for everything around us.

Now consider what we are trying to create.

Systems that can perceive.

Systems that can reason.

Systems that can learn.

Systems that can increasingly act.

And perhaps, eventually, systems that become more capable than humans in many domains.

That creates an interesting question.

What happens when the most intelligent species on Earth creates another highly capable intelligence?

The answer could be incredibly positive.

Such systems could become some of the greatest tools humanity has ever created.

They could help scientists discover medicines.

They could assist elderly people.

They could help people with disabilities live more independently.

They could improve education.

They could accelerate scientific discovery.

And in the case of the project I am working on, even something as seemingly ordinary as helping a visually impaired person walk confidently through an unfamiliar environment could have enormous impact.

But capability alone does not guarantee that every outcome will be good.

Humans themselves are evidence of that.

Intelligence gives an agent the ability to achieve objectives more effectively. The important question then becomes:

What objectives are we giving it?

And what happens when those objectives interact with a complicated world?

The Irony

There is something beautifully ironic about all of this.

Human beings spent thousands of years trying to understand nature.

Then we built machines.

Then computers.

Then software.

Then systems that could learn.

And now, to make those systems more intelligent, we are studying ourselves.

Our eyes inspire computer vision.

Our neurons inspired neural networks.

Our memory inspires artificial memory systems.

Our attention mechanisms inspired ideas in modern AI.

Our language became the training material for language models.

Our learning behavior inspires reinforcement and self-supervised learning.

Our interaction with the physical world drives research in robotics and embodied intelligence.

In some sense, one of the most advanced engineering projects humanity has ever attempted is reverse-engineering abilities that biology already performs every second.

And after spending hours thinking about neural networks, perception systems, spatial reasoning and intelligent agents, I can go outside and watch a five-year-old run across a playground while recognizing friends, avoiding obstacles, maintaining balance, understanding speech and deciding where to go next.

No server rack.

No giant GPU cluster.

Just a biological system doing its thing.

That is humbling.

What Started as a Computer-Vision Project Became a Bigger Question

I started this project thinking:

How can I build an AI system that helps a visually impaired person understand the world around them?

Now I find myself thinking about something much broader:

How do humans understand the world around them in the first place?

And perhaps that is one of the most exciting things about working in artificial intelligence.

Sometimes you start by trying to teach a machine how to see.

And somewhere along the way, you begin realizing how extraordinary it is that we can see, learn, move, understand and adapt at all.