Author: Xu Shanshan, Editor: Li Motian
In May this year, the last batch of Model S/X rolled off the assembly line, marking the end of production and retirement. Just 46 days later, the old production line was completely dismantled to build the first mass production line for the Optimus humanoid robot.
Optimus is the key to Tesla's transformation into an AI company. Its birth was highly dramatic: not only does it share a factory with the Model 3/Y, but its "brain" directly ports the FSD (Full Self-Driving) algorithm.
Tesla has repeatedly emphasized that the FSD algorithm in its vehicles can be reused for robots, turning the heavy-asset burden of the automotive industry into a software advantage with marginal costs approaching zero, thereby driving up its valuation.
Cathie Wood recently publicly admitted her mistake, stating that the value of Optimus was not factored into her $2,600 target price model for 2029 at all [1]: "We have been too aggressive in autonomous driving, but in humanoid robots, we might have been too conservative."
Elon Musk himself has frequently complained that Wall Street undervalues Tesla, repeatedly defending Optimus during earnings calls, calling the third-generation product a "true general-purpose robot" and a "product that is genuinely useful in daily life" [2].
However, behind the cheerleading lies a somewhat awkward time lag. Tesla first proposed this concept as far back as five years ago, yet it has not reached mass production to this day.
Humanoid robots are the crystallization of the software and manufacturing industries, inheriting autonomous driving on one hand and rooted in precision manufacturing on the other. Theoretically, this should be the comfort zone for automakers. Yet, watching robotics companies gain momentum, Tesla, despite its initial overwhelming advantages, has actually slowed down.
Where Did Things Go Wrong?
Algorithm Porting
In 2024, Andrej Karpathy, now running his own venture, appeared on a podcast and inadvertently revealed unknown details about Tesla's development of Optimus.
Andrej Karpathy studied under Fei-Fei Li and was a founding member of OpenAI. After joining Tesla, he quickly built up the AI team and was a key figure in developing Autopilot and FSD. In 2022, Karpathy returned to OpenAI, but his genuine appreciation for his former employer led him to let slip in the heat of the moment [3]:
"An early version of Optimus once thought it was a car because it used the same computers, cameras, and algorithms running on the cars."
Porting the FSD algorithm for Optimus is not Tesla's exclusive trick; there are indeed commonalities between humanoid robot and autonomous driving technologies.
First, their essences are similar: both aim to enable machines to understand the physical world and interact dynamically with real-world environments.
Second, both follow the "perception-decision-control" chain. Sensors perceive environmental information, the brain thinks and commands based on "what it sees," and the limbs execute the final instructions. This is the prerequisite for algorithm porting.
External estimates suggest that Optimus reuses about 60% of the FSD algorithm. He Xiaopeng's answer is even more aggressive, claiming that the software commonality between humanoid robots and cars is as high as 90%.
In other words, the capital market's logic of valuing Tesla as a tech company is not an exaggeration. Seeing this, other automakers quickly took action, packaging and stuffing their own algorithms into robots.
This has led to an interesting phenomenon: a significant portion of the engineers on the Optimus project team come from the vehicle department. Currently, the most crowded track for startups in the automotive industry is undoubtedly humanoid robots.
Although the slogan "automakers understand robots best" is shouted loudly, the reality of development is somewhat different.
Whether it is XPeng, which has just released its humanoid robot IRON, or Tesla, which is gearing up to unveil its third-generation trump card product, the focus of their demonstrations is on hardware and mobility capabilities.
The advantage of automakers is very likely not in algorithms.
Hardware Advantages
Compared to the brain algorithm porting, the effects of which are not particularly stunning, the advantages of automakers likely stem more from hardware, specifically two major areas.
First is supply chain management capability. Although the parts for cars and humanoid robots are not highly interchangeable, their technical principles are largely the same, resulting in a high degree of supply chain overlap.
Besides legacy demands like sensors and batteries, taking the new demand for joints—which account for the highest proportion of the whole machine cost—as an example, motors, reducers, and ball screws are the three major components, with a large number of automotive suppliers involved:
Hengshuai makes motors, Zhongding makes reducers, Shuanglin makes ball screws, Sanhua Intelligent Controls makes modules, and there is also Schaeffler, which does everything. Leaderdrive, the only domestic harmonic reducer company in the Optimus supply chain, also inadvertently broke into the automotive supply chain.
At the same time, coordinating hundreds or even thousands of suppliers and managing production and inventory with refined precision are also strong suits of automakers.
This means automakers can systematically transfer their supply chain management experience to the robotics business. Massive automotive orders not only enable automakers to secure extremely low procurement prices but also allow them to deeply participate in product R&D, ensuring a cost advantage.
Second is manufacturing and production capability. The automobile can be counted as one of the most complex industrial products in human history. Organizing large-scale standardized production amidst this complexity is a capability in itself.
The same applies to robots. Building a prototype is easy, but whether it can be mass-produced and what the yield rate is will determine whether the robot's cost can be reduced to a reasonable range.
Compared to startups operating in a cottage industry model, automakers possess highly automated production lines, mature four major manufacturing processes, and stringent quality control systems. Even if the production lines cannot be directly moved over, the production experience can be fully reused.
However, in terms of software algorithms, it is likely difficult for automakers to pull ahead of native robotics companies.
Algorithm Differences
For humanoid robots to achieve "general-purpose" capabilities, hardware is the threshold, but the key lies in algorithms.
General-purpose corresponds to specialized. Traditional robots are dedicated to specific tasks because their algorithms highly rely on functional models; if the shape of a cup changes or a table is 5 centimeters higher, the system will crash. Over the past two years, the rapid development of VLA (Vision-Language-Action), world models, and reinforcement learning has given robots the confidence to transition towards general-purpose capabilities.
Coincidentally, the mainstream architecture for autonomous driving algorithms is also VLA (Vision-Language-Action large models) and world models, but they cannot be simply copy-pasted for robots.
The physical worlds faced by cars and robots are completely different, and the algorithmic goals also have fundamental differences.
No matter how complex autonomous driving algorithms are, they serve the goal of "moving from point A to point B." Sensors must recognize traffic lights and obstacles to make decisions on acceleration or braking. The challenges are nothing more than how to handle pedestrians darting out from blind spots or vehicles cutting in, whether the driving strategy is comfortable, and whether there will be sudden jerky braking.
Humanoid robots are different. Recognizing a water cup on a table is for the purpose of picking it up; seeing a messy room means the task is to tidy up clothes and clean. From the perspective of long-term scenario versatility, the challenge for humanoid robots lies more in how to be equally adept in the living room and the kitchen.
Understanding the physical world means being halfway to success for autonomous driving, but for robots, it is merely taking a small step.
Autonomous driving algorithms only need to know what constitutes good driving skills and avoid intentional collisions or friction with people and vehicles; whereas the value of humanoid robots is precisely reflected in the contact phase, requiring an understanding of physical common sense to manipulate everything.
Assume a scenario: picking up an egg on a table. The humanoid robot must first identify the egg, confirm its position, and control the force of its hand during grasping to prevent the egg from slipping and falling to the ground or being crushed by excessive force.
In this scenario, the brain first judges the smooth and fragile attributes of the egg and issues a grasping command. The cerebellum then adjusts the grip force in real time based on tactile feedback. This involves both multi-body dynamics and tribology, which are typically not priorities for autonomous driving algorithms.
To be precise, the acceleration and braking commands output by autonomous driving algorithms cannot be converted into the skills required by robots at all.
No matter how well the autonomous driving algorithm performs, it is at best a specialized robot system; if a humanoid robot only moves "from point A to point B," it is merely an autonomous driving system with excellent driving skills. True general-purpose capability lies in being "all-capable," meaning what it can do and how well it does it.
Speaking solely of this point, data is the biggest roadblock facing automakers.
Data Challenges
Humanoid robots require massive amounts of data to train algorithms, essentially to break Moravec's paradox: things that are easy for humans to do are difficult for AI, such as grabbing an egg or pinching tofu with bare hands.
A three-year-old child can grab an egg without breaking it not because they have studied mechanics or understand friction, but because biological evolution instincts tell them how much force to use.
Without considering pre-programming and teleoperation, understanding common sense of life and physical rules, and acquiring human instincts through learning are inherently beyond the scope of what robots can easily acquire.
Therefore, compared to single driving data, the data needed for humanoid robots must cover all aspects and be continuous. The brain can learn how to cook by watching videos, but how much force to use when holding a spatula and what speed to flip food ultimately require more learning and practice.
The problem is that autonomous driving data can be collected by sitting in an office waiting for transmissions from mass-produced fleets, whereas acquiring data for humanoid robots is significantly more difficult.
In the past, data collection mainly relied on teleoperation of real machines. Operators wore VR headsets and force-feedback gloves, synchronizing the robot's perspective for 1-on-1 hands-on teaching. The advantage is precise operational data; the problem is that it cannot be done without operators, and the data easily becomes invalid if the machine is changed, resulting in high costs.
In the past two years, the concept of "data is king" has emerged, and highly efficient embodiment-free data has become popular. Collection tools have also become increasingly lightweight, mainly focusing on UMI and EGO. One uses a handheld special gripper to record hand movements, while the other uses wearable devices to record first-person perspective operations, making it convenient to do two jobs for the price of one [4][5].
In this way, embodiment-free data can be used for pre-training algorithm models, while real-machine data retreats behind the scenes for model fine-tuning, thereby improving overall efficiency.
A common sight in Indian factories: workers collecting data for almost zero pay while worrying about being monitored
In this data phase, the challenges faced by automakers and native robotics companies are basically the same. They can only take on the dirty and tiring jobs, with no difference between domestic and international markets.
Returning to automakers, the experience of building cars may help build robots, but whether the job is done well ultimately depends on data and algorithms. In this brand-new race for humanoid robots, I'm afraid no one holds the master key to an outright victory.
References
[1] Cathie Wood: Tesla Stock Forecast, Bitcoin Bull Case, AI Outlook, Morningstar
[2] Tesla Q4 2025 and Q2 2026 Earnings Calls
[3] With Andrej Karpathy from OpenAI and Tesla, No Priors
[4] ‘Who is going to pay us when we’re replaced by robots?’ The Indian factory workers told to film themselves for AI, The Guardian
[5] Indian Workers Wear Head Cameras to Teach Robots While Working, Taking Their Own Jobs, Cool Play Labs (coollabs)