For the past decade, capital has chased after “autonomous driving”. In the next decade, capital will bet on “humanoid robots”.

Recently, Fei-Fei Li said something in an interview that has been repeatedly quoted throughout the industry:

“Autonomous driving is two-dimensional, its goal is to avoid collisions; humanoid robots are three-dimensional, they need to interact with the world, manipulate the world, and understand the world.”

This statement seems understated, but it actually hits the essence of the humanoid robot industry:

While autonomous driving has not yet fully achieved L4, humanoid robots must operate in a more complex three-dimensional world—this is not just a little harder, but an order of magnitude more difficult.

1. Why is Autonomous Driving Only Two-Dimensional?

The ultimate goal of autonomous driving is fundamentally one:

To avoid collisions.

This means that the world of vehicles is a relatively abstract two-dimensional plane:

• Few objects

• Fixed rules

• Planned trajectories

• Predictable speeds

• Simple physical structure (road + vehicle)

In simple terms:

The ability to avoid obstacles is the most critical test for autonomous driving.

Is this task difficult? Yes, it is very difficult. But it has clear boundaries, mathematical models, and a controllable environment.

However— ten years have passed, and no one in the world has achieved stable L4 fully autonomous driving in all cities.

If the “two-dimensional obstacle avoidance problem” has not been completely solved, you can imagine the magnitude of the challenges faced by “humanoid robots”.

2. Humanoid Robots: From Two-Dimensional to Three-Dimensional + Interaction + Manipulation + Collaboration

Fei-Fei Li emphasized:

Robots need to interact with the world, which is much more difficult than vehicle obstacle avoidance.

Why?

Because robots do not simply “avoid”; they need to “intervene” in the world.

▍Autonomous Driving:

• Recognize → Avoid Obstacles → Move (just don’t hit others)

▍Humanoid Robots:

What they need to achieve is—

• Grasp a doorknob

• Open a bottle cap

• Pick up an egg without breaking it

• Fold clothes

• Go up and down stairs

• Carry a tray

• Use a screwdriver

• Interact with humans

• Understand “Will this action hurt someone?”

These actions involve:

Complex contact relationships + Three-dimensional movement + Force control + Perception + Intent understanding

These are abilities that humans have evolved over millions of years.

Cars do not have “hands”, nor do they have “interactions”, and they certainly do not need to “operate tools”.

Robots must possess all of these capabilities.

That is why Fei-Fei Li emphasizes:

“For robots to apply appropriate forces to objects in a three-dimensional world is the most challenging problem in artificial intelligence.”

3. The Hand is the Summit of Humanoid Robots

Breaking it down further, you will find:

To enable robots to do anything, the key is not AI, not vision, not walking algorithms, but—

The Hand.

The human hand has:

• 27 degrees of freedom

• Hundreds of tendons

• Complex skeletal interactions

• Tactile feedback

• Force distribution

• Dynamic adjustments

• Exceptional operational precision

And what about robotic hands?

Today, the best “dexterous hand” in the world (Shadow Hand) only simulates about:

60–70% of human operational capability.

Moreover, each one costs over a hundred thousand dollars, making them expensive and fragile, completely unsuitable for mass production.

So how is the industry currently approaching this?

There are three main routes:

1. Tendon-driven: Most dexterous, but the most expensive (research route)

2. Lead screw (direct push) driven: Most stable, maximum force, suitable for mass production (Tesla route)

3. Servo motor joints: Cheapest, easiest to mass produce (Unitree / Digit)

Each has its pros and cons, but there is still a long way to go before reaching “human hands”.

In other words:

We can make robots “walk”, but we are still an era away from making robots “do things with their hands”.

4. Why is Capital So Enthusiastic? Because Humanoid Robots Are More “Attractive” Than Autonomous Driving

The market for autonomous driving is the transportation industry. The market for humanoid robots is the “global workforce”.

McKinsey estimates: 350 million jobs worldwide could be replaced or assisted by robots, which represents a trillion-dollar industry.

No matter how large the autonomous driving market is, it cannot compare to “the workforce itself”.

Moreover:

• Large models have given robots a “brain prototype”

• Tesla, UBTECH, and Figure have received massive funding

• The supply chain (motors, reducers, lead screws, force control) is rapidly maturing

• Video models (Sora / Vidu) enhance the “teachability” of robots in understanding scenes

Capital sees that:

The end of autonomous driving is the humanoid robot in three-dimensional space.

Driving is a two-dimensional task. Working is a three-dimensional task.

The ultimate terminal of the entire AI system should not just be a car, but rather a “capable body”.

5. Timeline for the Implementation of Humanoid Robots

Based on industry consensus + Fei-Fei Li’s views + Tesla/UBTECH’s R&D pace:

▍2025–2027: Robots will take on simple repetitive tasks

• Handling

• Boxing

• Sorting

• Warehouse inspections

• Simple assembly

→ Essentially replacing “manual labor”.

▍2028–2032: Entering household scenarios (limited)

• Assisting with plating

• Loading the dishwasher

• Simple vegetable preparation

• Basic housework

• Assisting in elderly care

→ Achieving 40–50% of human operational capability

▍2033–2038: The emergence of true “humanoid labor”

• Skilled use of tools

• Highly stable grasping and manipulation

• Complex assembly tasks

• Human-robot collaboration

→ Considered the best carrier for AGI

▍After 2040: Full social integration (optimistic version)

Replacing a large number of basic labor positions.

6. Where is the True Value of Humanoid Robots?

You will find that this track is very similar to mobile phones and electric vehicles:

The biggest commercial barrier is not AI, but the supply chain.

The five core components are:

1. Reducer (tendon + torque amplification)

2. Lead screw/tendon (linear retractable muscle)

3. Drive motor (power source)

4. Highly integrated joint modules

5. Force control + tactile sensors (nervous system)

These are areas where Chinese companies excel, and they represent the largest new blue ocean in the future industry market value.

In summary:

AI makes robots “smart”, the supply chain makes robots “capable of working”.

7. Conclusion: Humanity’s First Attempt to Materialize “Ability”

Humanoid robots are the first physical embodiment we give to AI.

The starting point of all human civilization on Earth comes from two things:

• The ability to walk

• The ability to use hands

Therefore, I have always believed:

Humanoid robots are not a product, but a “revolution of bodies”. They are not a replacement for humans, but an expansion of human boundaries.

Autonomous driving is the end of the two-dimensional era, humanoid robots are the beginning of the three-dimensional era.

In the next decade, we will see machines truly “enter the world” for the first time.

And standing on the eve of this transformation, we are witnessing—a new industrial revolution taking shape.