Why have autonomous driving professionals fallen out of favor in the embodied intelligence field?

The successful experience of the previous generation is becoming ineffective, and the new world is starting to screen for new capabilities.

In the past two years, the embodied intelligence industry has almost absorbed the talents, methodologies, and organizational experience of the autonomous driving industry all at once.

But this year, more and more practitioners have begun to realize that there may always be a gap between autonomous driving and embodied intelligence.

In the past six months, I've talked with many entrepreneurs in the field of embodied intelligence and people in the robotics industry about a question: Do you still prioritize candidates with autonomous driving backgrounds when recruiting?

The answers have reached a consensus.

The halo of autonomous driving is fading

Around the Spring Festival, some headhunters mentioned that some robotics companies have clearly preferred to recruit people with experience in the robotics industry first, and the interview probability of cross - industry candidates has dropped significantly.

Having an autonomous driving background is not necessarily a negative factor, but it is no longer as naturally attractive as it was in previous years.

Some time ago, a company that started doing embodied intelligence after splitting from a leading autonomous driving company directly wrote "No autonomous driving background required" in the job requirements when recruiting world model researchers.

What's more interesting is that when I interviewed the CEO of an industrial robotics company not long ago and talked about talent recruitment, he also mentioned that the team basically no longer considers recruiting people with a pure autonomous driving background.

I subconsciously asked back, "But aren't you also from the autonomous driving industry?"

He smiled and said, "Embodied intelligence is a new industry. If you really rely on old experience, it will basically fail."

He explained that the reason for not considering candidates with an autonomous driving background much when recruiting is that although autonomous driving and embodied intelligence seem similar on the surface, their underlying logics are different.

This statement sounds a bit counter - intuitive.

After all, in the past two years, many autonomous driving companies have actively embraced new AI paradigms, whether it's VLA, world models, end - to - end, or data closed - loops.

From a technical stack perspective, the two share a large number of underlying capabilities, and everyone naturally assumes that switching from autonomous driving to robotics is a dimensionality - reducing strike and a logical step.

But more and more people who have truly delved into the field are starting to admit that we initially saw the two as too similar. The complexity of problems in autonomous driving is actually much lower than that in embodied intelligence.

Autonomous driving is of course complex, but in essence, it solves problems in a highly constrained environment —

The roads are fixed, the traffic rules are clear, the vehicle forms are unified, and most variables are within a relatively definable framework.

Robots, on the other hand, face an open physical world and need to consider almost all possible interaction relationships between humans and the physical world.

Tightening screws, feeding materials, sorting, assembling, handling, organizing, each action may involve a different set of data, a different set of training, and a different set of deployment methods.

This may also be the most core difference between the two:

There is no concept of downstream tasks in autonomous driving. You can't build a car specifically for driving on the 5th Ring Road in Beijing, nor can you make a car only learn the road conditions on the Outer Ring Road in Shanghai. It pursues a unified and general solution that can adapt to all road scenarios.

The essence of a robot is a collection of tasks. Today, no single robot model can cover all industrial scenarios, let alone household scenarios. Pretraining, post - training, scenario fine - tuning, few - shot learning... This cumbersome process is exactly the norm in the robotics industry.

And this is the most difficult part for many people who have grown up in the autonomous driving system to adapt to.

Experience is becoming a liability

The autonomous driving industry has developed for more than a decade and has formed a set of extremely mature and well - verified methodologies.

This system has trained a large number of excellent engineers who understand mass production, engineering, system architecture, and data closed - loops. These capabilities are invaluable in any high - tech industry.

In fact, most of the top entrepreneurs in the embodied intelligence industry today still come from the autonomous driving field.

But the problem is not whether experience is valuable, but when experience will become a burden.

Engineers who have grown up in the autonomous driving system for a long time are prone to form a strong path dependence: they are used to looking for unified solutions, used to building general frameworks, and used to explaining and solving new problems with methods that have been proven successful in the past.

This leads to having a "heavy hammer" in hand, seeing everything as a "nail".

But embodied intelligence often does not reward this kind of thinking. To some extent, it may even punish it.

Because the biggest characteristic of the robotics industry today is that there are almost no standard answers to most things.

Many entrepreneurs have privately said similar things: When they evaluate people, they care most about neither whether the person has worked on robots nor whether they have worked on autonomous driving, but whether the person can quickly understand a brand - new problem.

An entrepreneur in the field of embodied intelligence once said something very interesting:

"If there are two people today, one with five years of autonomous driving experience and one who has just switched from AI research, I may not necessarily choose the former."

The reason is simple: An autonomous driving background means you have solved many definite problems in the past; while the robotics industry needs more the ability to dare to start over and reset when facing unknown problems.

This reminds me of what was popular in the autonomous driving industry when it first emerged a decade ago: Autonomous driving companies do not prioritize recruiting people with a vehicle engineering major.

Embodied intelligence needs a different kind of person

So, what kind of people are today's embodied intelligence companies really competing for? Based on the views of dozens of industry insiders over the past period, I've summarized four core criteria:

First: Have experience in AI.

Here, it doesn't necessarily mean having worked on robots or autonomous driving, but truly understanding how models, data, training, and inference work, knowing how to quickly verify an idea, and knowing how to turn an abstract judgment into an experiment.

Second: Have no heavy historical baggage.

When seeing new things, the first reaction should not be "How I did it before", "How Waymo did it", or "How Tesla did it", but to first judge whether this path makes sense at present. The embodied intelligence industry has not yet formed a unified paradigm, and many problems are still in the definition stage. Bringing in too much existing experience may easily lead you astray by old answers.

Third: Have solid mathematical and abstract thinking abilities.

Because many problems in the field of robotics today have no ready - made answers, in the end, it still comes down to modeling, abstraction, and engineering implementation.

Fourth: And most importantly, be willing to accept new things and be willing to learn anew. This has almost become the core criterion for many CEOs of embodied intelligence companies when evaluating people. Because in this industry, the truly scarce ones are those who can best adapt to new variables.

This also explains why many embodied intelligence companies are still absorbing talents from the autonomous driving field while also being more vigilant about candidates with an autonomous driving background.

They don't reject people from the autonomous driving industry, but they do reject those who are too bound by the autonomous driving methodology.

Of course, we can't deny the huge contribution that autonomous driving has brought to the embodied intelligence industry. It has not only provided the first batch of talents and funds, but more importantly, it has verified the feasibility of the path of AI + the physical world.

But the complexity of autonomous driving and embodied intelligence is, to a large extent, ultimately incomparable.

The difficulties in autonomous driving lie more in decision - making, and the control system has become quite mature after decades of development;

While the control of robots has not been completely solved yet, the decision - making space is several orders of magnitude larger than that of autonomous driving, and coupled with the still - uncrossed gap from language to action, the difficulty increases exponentially.

Autonomous driving needs to reach nearly 100 points to be legally deployed; even one point less may lead to a major disaster; while embodied intelligence may have a chance to enter the scene with 60 points or even 40 points —

The risk of a household robot accidentally breaking a cup is completely different from the accident risk of an autonomous vehicle.

Autonomous driving has a perfect transitional stage of L2, where companies can make money while iterating; while embodied intelligence has no intermediate state. It can either truly replace humans in completing tasks (L4) or be just a toy (L0).

Frankly speaking, most of the lessons learned in autonomous driving entrepreneurship will probably have to be learned again in embodied intelligence entrepreneurship. Even if a few people can see further by standing on the shoulders of giants, most people will still fall into the same pit.

Today, embodied intelligence has achieved some stage breakthroughs in the robot learning paradigm, but it is still far from systematically replacing human efficiency in the physical world like autonomous driving. That path will be a long one.

And for individuals, the choice has become extremely realistic.

Embodied intelligence offers higher salaries and stands at the forefront of the capital - favored trend, and its ecological niche is rapidly taking shape. This suddenly reminds me of a conversation I had with an engineer who recently switched from autonomous driving to embodied intelligence. He half - jokingly said, "I'll make some quick money here for a few years. If the embodied intelligence industry doesn't work out, I'll just switch back."

"The positions are filling up. Get in early."

This article is from the WeChat official account "ZhiJian autoweek", author: Wang Asen. Republished by 36Kr with permission.