机器人“大脑”60年进化史：基础模型的五代进化与三大闭源流派｜机器人特辑

The demos released by robotics companies in 2025 are all somewhat fantastical, starting with the figures. In October, AI released its third-generation robot capable of performing various household chores.

The demo was quite impressive , but the success rate of the tasks was met with considerable skepticism within the industry.

I also want to criticize the design ; its face exhibits a rather pronounced uncanny valley effect .

Meanwhile, another prominent company, 1X, released a demo at the end of October that was significantly smarter, with a much cuter facial design that seemed more appealing.

However, this robot, called Neo, relied on remote control and was criticized as "pseudo-intelligence," with various privacy concerns.

Tesla's robot , while also releasing various demos , including a very smooth running demo in December , clearly faces significant challenges to its 2025 mass production plan , forcing the company to halt production and redesign the hardware.

Furthermore, its most recent appearance encountered a minor mishap.

What was it?

We'll keep you in suspense for now.

Our robot series video has already discussed this, and we also released a comprehensive summary of embodied intelligence in 2025.

Next, we'll delve into a core technology of this industry: the basic robot model.

We're attempting to answer the question: why has 2025 suddenly become the "Year Zero" for basic robot models? We

also visited cutting-edge robotics companies and labs in Silicon Valley.

The Basic Model section will be divided into two episodes: "Closed Source" and "Open Source."

We'll systematically dissect how the "brains" of today's mainstream robots are trained , how they connect to the real world , and the technology and business logic behind different approaches.

We'll show you how the robot brains of the large-scale model era truly develop.

Yours isn't even as aesthetically pleasing as a stack of machines!

In this episode, we'll first discuss the current darling of the capital market: closed-source systems. But before answering these questions, we need to understand a more fundamental concept: what exactly is a robot basic model?

If we were to explain robot basic models in one sentence...

The simplest analogy is that if GPT is a "talking brain," then the basic robot model is a "hands-on brain." However,

it took humanity a full 60 years to create this "hands-on brain."

Let's first review the four major robot paradigms before the advent of the large model.

Let me take you back to 1961, the year the world's first industrial robot, Unimate, "went to work" in a General Motors factory.

Its job was simple: pick up hot metal parts from one production line and place them on another.

From today's perspective, it seems quite "silly" because it relied entirely on programming engineers to tell it by code: Step 1: Move the arm 30 centimeters to the left; Step 2: Close the gripper; Step 3: Move the arm upward 50 centimeters; Step 4: Rotate the arm 90 degrees to the right; Step 5: ... Releasing the gripper

Step 5: ... Releasing the gripper sounds silly, right?

But at the time, it was a revolutionary breakthrough.

The problem with this method was obvious: zero tolerance for error and zero flexibility.

If the part was off-center by 1 centimeter, the robot couldn't grasp it.

If a different sized part was used, the code had to be rewritten.

Not to mention dealing with unexpected situations , such as a part falling to the ground, the robot was completely lost.

In the highly controlled environment of a factory, this method worked for decades , and even today, many automotive welding robots still use this "programmable" logic.

But in the 1990s, roboticists realized that programming alone wasn't enough; robots needed to be able to "perceive" their environment.

This led to the development of SLAM (Simultaneous Localization and Mapping) and motion planning technologies .

The core idea here is to first use sensors to "see" the surrounding environment , creating a 3D map , then plan a path on the map, and finally execute the action.

Guess what the most successful application of this method is?

The once-popular Roomba robotic vacuum cleaner works this way.

It uses LiDAR to scan the room, create a map , plan a path covering the entire area, and then moves along the path, avoiding obstacles.

This method is crucial for "guiding" the robot's movement.

While successful in "perception" tasks, early unmanned vehicles, drones, and logistics robots largely followed this approach.

However, it failed in "operation" tasks.

Why?

Because operation tasks are too complex.

Imagine a robot folding a towel.

The traditional method involves four steps: first, visually identifying the four corners of the towel; second, calculating the 3D coordinates of each corner; third, planning the arm's movement trajectory; and fourth, executing the grasping, folding, and placing.

This sounds reasonable , but in practice, it 's fraught with problems. The towel might become crumpled, making it impossible to identify the "four corners."

Furthermore, towels are flexible ; a single grasp deforms them, immediately invalidating the 3D coordinates.

Every step is prone to error, and an error can cause the entire process to collapse.

In 2010, a research team at UC Berkeley conducted an experiment using this "perception → planning → execution" method to fold towels . The result? An average of 24 minutes was taken to fold a towel.

Even in today's AI era, towel folding remains a crucial task requiring a foundational model to drive the robot's success—a topic we'll discuss later.

Since manually designing rules is ineffective , can we allow robots to directly "learn" how humans do it?

This is called behavior cloning. The idea behind cloning, also known as imitation learning, uses the example of folding a towel . Robots learn by imitation

. Robots learn by imitation by having a human demonstrate how to fold a towel many times , recording the visual input and action output for each frame , training a neural network to learn the mapping between input and output, and finally, when the robot sees a towel, it directly outputs what action to take.

(2015, Google...) A team at Brain used this method to teach robots to grasp various objects.

They collected hundreds of thousands of grasping data points to train a neural network , pushing forward the advancement of "vision-action" learning in robotic grasping tasks.

This was a huge leap forward; for the first time, robots could learn from data without manually writing rules.

However, this method had a fatal flaw : its data efficiency was too low.

It required hundreds of thousands of grasping data points for training, and this was just for the single action of "grabbing."

To learn "folding towels," even a million demonstrations might not be enough.

More critically, this method had very poor generalization ability.

A model trained on data collected from model A robots would basically be unusable on model B robots.

Then came the reinforcement learning era in 2010.

In 2016, AlphaGo defeated Lee Sedol, proving the power of reinforcement learning.

Robot scientists were also wondering if robots could also use reinforcement learning to figure out how to complete tasks on their own.

The core idea of ​​reinforcement learning is that it does not require human demonstrations.

The idea was to let the robot learn by doing, rewarding it for correct mistakes and punishing it for incorrect ones.

The robot would gradually learn how to maximize its rewards.

Boston Dynamics began incorporating reinforcement learning into its motion control systems, enabling robots to walk, jump, and perform backflips on complex terrains.

However, reinforcement learning had a major problem : it was too slow.

AlphaGo played tens of millions of games with itself in a simulation environment to learn Go , but robot operation tasks are difficult to practice in a simulation environment because the environment is too complex, difficult to set up , and differs significantly from the real physical world, leading to inaccurate simulations.

But real-world testing was too slow, too expensive, and too dangerous.

Imagine teaching a robot to fold a towel ; it might have to try millions of times.

Most of the time, it would miss , drop the towel, tear it , or get its arm stuck, etc. At this rate, it would take forever to learn.

Moreover, reinforcement learning has a more fundamental problem : it doesn't know "constant..." Humans know that towels are soft, foldable, and have some friction , but reinforcement learning robots need countless trials and errors to "discover" these common-sense facts —meaning the efficiency is too low.

The story finally came to 2025 with the emergence of large language models, which changed everything.

In 2022, ChatGPT burst onto the scene.

People discovered that large language models contain a wealth of "common sense" about the human world.

It knows what a towel is, what folding means, what to do first, and what to do next.

It has reasoning, planning, and generalization abilities.

The industry's first reaction was whether large language models could be combined with robots, and thus the VLA (Vision-Language-Action) model was born.

The revolutionary aspect of the VLA model lies in its unification of three things into a single neural network : Vision ( seeing the current scene ) and Language (language). Understanding task objectives, common sense , and actions, VLA outputs specific control commands.

For example, if you tell a robot, "Put the apples on the table into the basket," the traditional method requires four steps: first, visual recognition of "apples" and "basket" ; second, planning the trajectory for "grabbing the apples"; third, planning the trajectory for "moving to the basket"; and fourth, planning the action for "putting down. "

VLA, on the other hand, is an end-to-end neural network that directly outputs "what action to take next" from "language commands + visual input.

" Even more amazingly, it can perform "common sense reasoning."

For example, if you say, "Prepare breakfast for me," it knows from its home environment to take eggs out of the refrigerator , handle them carefully to avoid breaking them, and put bread in the toaster.

These common sense principles don't require you to program them line by line , nor do they need to "discover" them through millions of trial and error because they are already present in the large language model. In

fact, at the architecture level, we use VLA . Simply put , VLA

. Simply put , VLA uses the large domain model VLM as the backbone , but we transform the final output into actions usable in the robotics domain.

Intuitively, an Action can be translated into commands like "I want to move this arm to a certain coordinate point.

" The biggest criticism of VLA is why we need it at all.

Language is crucial because many traditional robotic algorithms are purely vision-based, and humans also rely heavily on vision when performing tasks.

However, if you think about it carefully, your brain actually generates something similar to language to tell you what to do first and second in a long-term task.

For example, when you're trying to unscrew a bottle cap, you don't mentally think of the words "unscrew the bottle cap" before you actually do it, because the language model is already quite small.

So, we believe the role of language models (LLA) is to better combine language and vision for very complex tasks. They can learn a lot of logic by training on a large language model.

For example, if you want to drink water, you need to find a cup or a bottle, right?

This is something that a large language model can directly provide.

The main purpose of using VLA is to better combine language and vision.

Otherwise, if you only have vision, the tasks you can do are likely to be short-term; you can't do any long-term tasks that require reasoning.

This is the main reason why we focus so much on introducing language.

So why did 2025 become the "first year of embodied intelligence foundational models"?

Because three key factors matured simultaneously in that year.

The first factor is that large language models are "sufficient."

From 2024 to 2025, companies like OpenAI, Anthropic, and Google successively released new models.

Large language models are now nearly "mature," at least sufficient for the capabilities required by robots —understanding instructions, planning tasks, and common-sense reasoning.

Firstly, the large models themselves are approaching maturity.

You can see that recently, whether it's OpenAI or other companies , the models they release are showing incremental growth , not the leapfrog growth seen when it went from 3.5 to 4.

Therefore, we believe that the capabilities of large models have stabilized and are sufficient to provide a good foundation for embodied intelligence . This is the most important factor from the model level.

. This is the most important factor from the model level.

Secondly, computing power prices have been halved again; renting an NVIDIA H100 in 2023... GPUs are

incredibly expensive and require long waiting lists to obtain.

However, with the price war among GPU cloud service providers and the mass availability of NVIDIA GPUs, startups can now afford to rent thousands of cards to train models.

The second factor is the increasing overall computing power.

Every year, as other chip companies develop more powerful chips, the equivalent computing power price decreases. In

a few years, the equivalent price might be half of what it was. Therefore, the enhancement of computing power has a significant impact

it was. Therefore, the enhancement of computing power has a significant impact on embodied intelligence.

The third factor is the maturation of the hardware supply chain.

Many people haven't noticed this change.

In 2024, with the surge in humanoid robots, a large amount of capital flowed into upstream component manufacturers , especially Chinese suppliers.

Motors, reducers, and sensors were originally niche products , but since 2024, several suppliers have received substantial funding and started expanding production.

Lower hardware prices have reduced the barrier to entry for robot manufacturing.

Finally, the maturity of various robot hardware components is relatively high.

Especially since last year's surge in humanoid robots, many have invested significant effort and capital in basic components, including motors.

The maturity and cost of R&D for components such as reducers have both improved and decreased, so we feel that the timing is more opportune.

These three key elements have made 2025 a special window of opportunity, and a new generation of robots based on VLA will emerge.

2025 is the year of humanoid robots.

The first robot butler has finally made its debut.

Humanoid robots are expected to become one of the largest industries in history, a market worth 5 trillion, with 1 billion robots worldwide.

However, the VLA model is not perfect, and its core challenge is the large amount of data. Language models can be trained with text from the Internet , but robots need "real machine data," that is, sensor data from the robot itself.

This kind of data simply doesn't exist on the Internet.

There are countless videos of people folding clothes on YouTube , but none of them are real machine data.

The video shows you the angle of your finger joints and the amount of force you apply when folding clothes.

This is why the core of this "arms race" is not only algorithms but also data.

Whoever can collect the highest quality data at the lowest cost will dominate the market.

We will also have a separate episode about robot data, so don't forget to follow our channel and don't miss the updates!

So, you see, the basic robot model didn't just appear out of thin air; it's the culmination of 60 years of technological accumulation.

It has accumulated the "precise control" of programmable robots, the "environmental perception" of model-based methods, the "demonstration learning" of behavior cloning , the "self-optimization" of reinforcement learning, plus the "common sense reasoning" of large language models.

This is the real "basic model."

You might want to know what level robots equipped with VLA models have reached now . We also visited Dyna Robotics,

. We also visited Dyna Robotics, a hot robotics company in Silicon Valley.

Among its three Chinese founders, Lindon Gao and York Yang are serial entrepreneurs.

Their previous AI shopping cart company, Caper AI, was acquired by Instacart for $350 million.

Jason... Ma is a former DeepMind research scientist specializing in basic robot models.

This company, founded just one year ago , has already completed two rounds of financing: a $23.5 million seed round in March 2025 and a $120 million Series A round in September of the same year, valuing the company at over $600 million.

The list of investors is truly impressive, including Nvidia, Amazon, Samsung, LG, and others . What first made them famous

. What first made them famous wasn't some fancy task or demo , but rather the very simple tasks of "folding towels" and "folding clothes."

On a whim, I suddenly wanted to have a speed contest with the robot and York to see who could fold clothes better and faster . Okay, let's start the timed competition.

. Okay, let's start the timed competition.

Although we're both faster than the robot , to be honest , I'm not sure... Dyna's robots fold towels well , and the key point is that although they are currently relatively slow , they can operate 24/7 without rest.

As long as the economics are sound, implementation is feasible.

In many scenarios where a person might work 8 hours a day, a robot can run 15 or 24 hours a day , effectively compensating for some of the efficiency loss.

Secondly, folding (towels) is indeed a good commercial application scenario because it is relatively simple and a fixed task.

However, in countries with high labor costs like the US, a significant amount of money would be spent on this.

Therefore, the businesses and clients we discussed are very willing to use robots for this task . After the competition,

. After the competition, let's talk about the main schools of thought in closed-source model robots.

Having looked at Dyna Robotics , let's examine the players in the robotics field in 2025.

We can divide them into three schools.

On the surface, they compete on technology routes, market share, and funding valuations , but essentially they are all vying for the same question : what is the correct path to achieving a "general-purpose robot"?

The first school is represented by Tesla Optimus and Figure AI.

The core belief of this school is that the basic robot model cannot be separated from the hardware; it must be vertically integrated and deeply coupled to achieve maximum effectiveness.

Tesla is the most radical representative of this school . Musk once said something quite audacious

. Musk once said something quite audacious : "80% of Tesla's value will come from Optimus robots."

Where does Musk's confidence come from?

Perhaps it comes from Tesla's ten years of accumulated experience in FSD (Full Self-Driving).

Milan, the former engineering director of Tesla Optimus... Kovac

once said, "We've just gone from robots on wheels to robots with legs."

Tesla has real-world data collected from millions of vehicles, an end-to-end neural network architecture, and a massive annotation team , so his logic sounds impeccable.

If FSD allows cars to drive autonomously in complex road conditions, why can't the same architecture enable robots to operate autonomously in complex environments ?

Both are closed loops of perception, decision-making, and execution; both are end-to-end neural networks ; the only difference is that the output has changed from "steering wheel angle" to "joint angle." However, the reality of 2025 wasn't so rosy.

At the beginning of the year, Musk confidently stated in an internal meeting that 5,000 Optimus robots would be produced by 2025, with 1,000 deployed in Tesla's own factories.

But by mid-year, after assembling just over 1,000, Tesla's Optimus humanoid robot production plan had been suspended, facing redesign.

And the recent controversy surrounding Optimus, mentioned earlier , occurred when it suddenly...

it suddenly...

The robot made a gesture as if it were trying to pull something nonexistent off its head and then tripped and fell.

This action was so similar to a human operator removing a headset that the video quickly went viral.

Many people immediately questioned whether Optimus was being remotely controlled by an operator.

Optimus's development doesn't seem as smooth as Musk claims. Where does the problem lie?

They were one of the earliest companies to create a humanoid robot and had some demo demonstrations.

They mainly use human video for transfer.

Its advantage is undeniable: human video is the easiest to collect because it doesn't require any external devices , and the collected data is from human hands operating the robot.

In terms of scalability, Tesla's model has the highest scalability.

However, its main problems are: first, the human hand and the robot's hand need to be very similar if you want the ability to transfer well.

This is why many people are now working on very dexterous hands that are very close to human freedom.

This is inherently a very difficult thing.

Secondly, even if you get closer, it's not exactly the same.

So there's still a gap between robot data and human data, what we call the "embodiment gap."

This embodiment gap is widely recognized in both academia and industry as a difficult problem to solve , resulting in low data transfer efficiency.

Tesla wants to train its basic model using massive amounts of human videos , but the physical differences between human and robotic hands —the so-called embodiment gap— are an insurmountable hurdle.

Even if you have all the human operation videos on YouTube, conversion efficiency is still an issue.

This is the first dilemma of the full-stack integration approach: you control the entire chain , but that also means every link in the chain is your bottleneck. If

the hardware isn't good enough , even the strongest model is useless; if the model isn't strong enough, even the best hardware won't perform to its full potential.

But Tesla has an advantage : plenty of money, a large workforce, and Musk.

These should be three advantages.

Will Optimus ultimately succeed?

We may have to wait another two years to see.

Figure AI takes a similar but more aggressive approach.

This company was founded in 2022.

Founder Brett Adcock previously worked on electric vertical takeoff and landing aircraft, essentially crossing over from "flying robots" to "ground robots."

Figure AI is expected to launch in early 2024. Figure AI made a bold decision to collaborate deeply with OpenAI, directly integrating GPT-4 into humanoid robots.

During that period, their demos were groundbreaking, seemingly capable of understanding human commands conversing and making independent decisions.

For example, if you asked, "Can I have some food?"

it would proactively offer you an apple.

However, this success was short-lived.

In February 2025, Figure AI suddenly announced its separation from OpenAI, declaring its intention to independently develop its own basic model, no longer relying on OpenAI's technology.

But just two weeks later, Figure AI swiftly launched the new Helix model, positioned as a general-purpose humanoid robot (VLA model) , emphasizing its completely self-developed control system for the entire humanoid robot.

It must be said that Figure AI's ability to forgo OpenAI's "crucial support" demonstrates its capabilities . Helix innovatively employs a System 1 and System 2

. Helix innovatively employs a System 1 and System 2 dual-system architecture.

System 2, like the cerebral cortex, is responsible for "thinking clearly about what to do.

" 1. Just like your cerebellum is responsible for "how your hands and feet coordinate."

When you pick up a cup to drink water, the cerebral cortex only needs to decide "now it's time to pick up the cup," and the cerebellum will automatically mobilize more than 20 muscles to complete the grasping action.

You don't even need to be aware of it.

This architecture solves a long-standing problem that has plagued robots.

Visual-language models are smart but too slow.

Traditional robot control strategies are fast but not versatile enough.

Helix allows both to perform their respective functions and trains end-to-end.

It can understand complex instructions and control them precisely in real time.

What's even cooler is that Helix uses a single neural network to control 35 degrees of freedom of the entire upper body, including the wrist, torso, head , and every finger.

It can also control two robots to work together to complete tasks at the same time.

This is the answer sheet that Figure handed in after "parting ways" with OpenAI.

September 2025 Figure AI completed a $1 billion Series C funding round, with its valuation soaring from $2.6 billion to $39 billion— a 15-fold increase in less than a year.

The list of investors reads like the Oscars of the tech world: Microsoft OpenAI Nvidia Bezos Intel Samsung etc. , already making them the "big shots" in the embodied robotics field.

To summarize, the core idea of ​​this school is that the generality of the basic model comes from being "large enough and end-to-end enough."

As long as there are enough model parameters, enough training data, and enough deep hardware and software integration, emergent capabilities will naturally emerge.

This is an experience summarized from the success of GPT-4.

However, whether this experience holds true in the physical world is still unknown.

If the full-stack integration school pursues "one-step generality," then the vertical breakthrough school pursues "emergence from specialization to generalization."

Their core idea is that instead of training a large model that can do everything but does nothing well, it's better to first perfect the model in a specific vertical scenario . The "learning ability" accumulated in this process

. The "learning ability" accumulated in this process will naturally transfer to other scenarios . Robotics

. Robotics is a clear representative of this school of thought.

We just saw the demonstration of Dyna robots.

Dyna takes a unique approach, creating general-purpose robots , but at the model level, they first utilize mature capabilities to implement them in practical scenarios to understand industry know-how and better guide algorithm research .

In other words, they first let the robots "work" in scenarios like laundries, restaurants, and gyms, learning while doing work. In April 2025,

work. In April 2025, they released "the first robot foundation model that can continuously operate at high performance in real-world environments, " DYNA-1. In 24 hours,

" DYNA-1. In 24 hours, their robot autonomously folded more than 700 napkins with a success rate of over 99.4% , requiring no human intervention and achieving a throughput of 60% of human speed.

But clearly, Dyna's ambitions go beyond just folding towels.

First, we want to clarify that we are not a company that makes towels.

Our Foundation Model contains all kinds of data, including various types of folding, such as folding towels, napkins, and clothes , as well as things like chopping vegetables, cutting fruit, preparing food, making breakfast, cleaning, or even sorting in logistics scenarios.

In fact, we have all kinds of data in our Foundation Model. The model (base model) is actually a very broad model.

Our generalization ability mainly comes from the base model.

We hope that the base model has strong enough capabilities to not require much customization for most tasks.

In the early stages, you may find that the process of migrating to a new task is relatively redundant and complicated . You will need to collect a large amount of new data

. You will need to collect a large amount of new data and mix it together for training.

However, as the amount of data in your base model increases, you will find that even when migrating to a completely new task , the migration cost will decrease . In the past, it might have taken several months of data

. In the past, it might have taken several months of data to migrate to a certain task , but now some simple tasks can be migrated with only one or two days of data.

So overall, as long as your base model becomes stronger and stronger , that is, as I just mentioned, its learning ability becomes stronger, then the model can be migrated to a new task.

The ability to transfer to new tasks also becomes stronger.

Dyna's understanding of the base model is completely different from that of the full-stack integration approach . Their understanding of the base model is that rather than training a generalized model that can do everything but does nothing well, it is better to let the model deeply specialize in a specific task first.

The "learning ability" accumulated in this process will help it master other tasks faster.

Just like someone who has mastered the piano to a music conservatory level will learn the guitar much faster than someone who has never learned an instrument at all, because they have mastered the meta-skill of "how to learn".

We have indeed seen that when your ability in a single task is greatly improved, it has a certain promoting effect on learning new tasks.

That is, if we use the highest quality data to do pre-training on the base dataset, the model will be able to extend to new tasks faster and require less data.

So this is also a counterpoint we have found in practice. Intuitive (counterintuitive)

, but it did happen.

Our understanding of it is more like that of a person: if you have a strong learning ability, I don't need to say what knowledge you have , but your learning ability itself is strong , and you will be strong in learning new abilities.

The premise of a strong learning ability is that you may have practiced and learned on many tasks in the past . Therefore, we believe that the learning ability itself and the learning process

. Therefore, we believe that the learning ability itself and the learning process are related.

This idea is based on the view that the scaling law of robot basic models and large language models may be different.

The rule of large language models is that the larger the model and the more data, the better the performance.

However, the performance bottleneck of robot basic models is not only in "model capacity" and "data volume", but also in "data quality" and "physical consistency".

If the physical contact in the training data is inaccurate, the model will learn the wrong physical intuition.

The larger the parameter, the more the error will be "amplified".

Why is it based on, to some extent , its scaling Law is certainly not as simple and straightforward as large language models.

In the field of language , they've discovered that even with a lot of low-quality data —like a string of text interspersed with ads— a relatively good model can still be trained.

This is because, with enough data, the model automatically filters out ads.

This is similar to humans ; after browsing many web pages , your eyes automatically filter out ads.

So, large language models can do this.

However, for robots, scaling law relies more on high-quality data .

As mentioned earlier, the data essentially consists of the robot itself plus video, which introduces a high degree of uncertainty . If you include a lot of

. If you include a lot of complex data, the robot model might not know where to focus its attention , resulting in less effective performance. What

we see now is that if the data quality is good enough, with increased data volume and diversity, the overall basic model capability will greatly improve , and the performance of downstream fine-tuning tasks will also be significantly enhanced.

This is something that can be tangibly observed.

Therefore, Dyna chooses a "small but precise" approach.

They believe that rather than training a generalization model with 10 billion parameters, it's better to train a... The first requirement for a specialized model with 1 billion parameters is ensuring that every training data point is a high-quality, realistic physical interaction . Secondly, it's crucial to enable the model

. Secondly, it's crucial to enable the model to self-optimize through reinforcement learning during actual deployment.

They believe that in the process of deep specialization in a task, the model learns not only "how to fold a towel" but also the meta-capability of "how to quickly learn new tasks."

Therefore, we are currently focusing on pathways like reinforcement learning and learning the basic capabilities of the large model, such as folding and arrangement abilities, to give it the ability to self-iterate and learn new skills .

I think this is the most important.

Ultimately, we believe that the basic large model should be able to handle ordinary tasks , such as household tasks like getting water or opening a door.

Similarly models that emphasize meta-learning capabilities include Skild. AI companies that start with "general-purpose models"

include Skild. AI companies that start with "general-purpose models" but don't manufacture hardware operate on the core logic of training a general "brain" using large-scale simulation data . This brain can then be quickly adapted to

. This brain can then be quickly adapted to different robot hardware and task scenarios .

For example, the same model can control a robotic arm to grasp objects , enable a quadruped robot to walk , and command a humanoid robot to perform complex operations.

Instead of training from scratch for each task, they rely on a powerful base model for transfer learning.

Rumors suggest that Nvidia and SoftBank will lead a $1 billion investment, valuing the company at $14 billion.

Another notable player in this field is Amazon.

In July 2025, Amazon announced the deployment of its 1 millionth robot.

What does 1 million mean?

Amazon currently has 1.56 million employees , meaning the number of robots will soon exceed the number of human employees.

However, these 1 million robots are all specialized, optimized for specific scenarios . Hercules can carry 1250 pounds of goods,

. Hercules can carry 1250 pounds of goods, and Pegasus is used for package sorting and transportation.

But Amazon's ambitions don't stop there.

Their Agentic AI team is developing a general-purpose robot base model and has even built a "humanoid" in their San Francisco office.

Amazon's strategy, similar to Dyna's, is to train humanoid robots to handle complex obstacles in an indoor testing ground called "park." Rather than creating a large, general-purpose model from the outset , Amazon focuses on accumulating the world's best data and strongest capabilities in vertical scenarios before generalizing.

If the first two camps are arguing about "whose path is faster," then the third camp is arguing about "who can set industry standards."

Their core belief is that in the basic model race, the ultimate winner isn't necessarily the one with the strongest technology , but rather the one with the strongest ecosystem control.

NVIDIA's logic is simple: to create the Android of robotics.

At the GTC conference in March 2025 , Jensen Huang prominently introduced the GR00T N1 and open-sourced it.

Sounds great, right?

But if you want to use the GR00T N1 , you have to use the entire NVIDIA ecosystem ; you can't escape it.

This is the power of ecosystem lock-in.

Once you use NVIDIA's entire toolchain, the switching cost is prohibitively high.

NVIDIA's moat isn't the model itself, but the entire ecosystem.

Google takes a different path: building influence through open-source research.

Google's robotics strategy has chosen a "research-driven, open-source-first" approach.

Its RT series emphasizes large-scale robot demonstration data and cross-task/cross-platform general models , and has established a strong influence in the academic and research community through papers and open datasets.

After the release of Gemini 3, Google has recently accelerated its pace and even poached former Boston Dynamics CTO Aaron Saunders to promote Gemini AL as a general robot control platform.

What are the essential differences between NVIDIA and Google's ecosystem strategies ? Are they truly open source or pseudo open source?

? Are they truly open source or pseudo open source?

We will discuss this later.

OpenAI and Meta are another type of player in this camp , taking small steps and moving quickly just to secure their position.

The relationship between OpenAI and robots is like a couple that has broken up and reunited . As early as 2018,

. As early as 2018, they made a breakthrough in robotic arm-operation tasks, but then the team size and priority shrank.

In 2024 and early 2025, they staged a drama with Figure, from a passionate relationship to a sudden breakup.

However, in the second half of 2025, they began to recruit more researchers focusing on humanoid robot control algorithms. In addition, OpenAI is also trying to build its ecosystem influence by investing heavily.

In November 2024, OpenAI and Jeff Bezos jointly participated in Physical Intelligence's $400 million funding round and Meta's strategy are similar but more low-key.

In early 2025, Meta established a new robotics division under its Reality Labs , led by former Cruise CEO Marc Whitten, with the goal of developing a humanoid robot platform.

Meta's CTO, Andrew Bosworth, has publicly stated that his team is building a "World Model " to support robots in performing more nuanced maneuvers than "walking" and "running and jumping."

On the surface, the three camps are vying for technology, market share, and valuation , but essentially they are betting on three common yet different assumptions about "generality."

about "generality." The full-stack integration camp believes that generality = a sufficiently large model + sufficient data + sufficiently deep hardware and software integration ; as long as these three conditions are met, emergent capabilities will naturally emerge —an experience learned from the success of GPT-4.

The vertical breakthrough camp believes that generality equals the transferability brought by deep specialization; robot scaling... Unlike language models,

robot scaling... Unlike language models, "small and precise" may be more effective than "large and comprehensive."

The key is to find the right "meta-learning" path.

Ecosystem platforms believe that universality equals the degree of ecosystem standardization.

Who wins in terms of technical route is not important; what matters is getting everyone to use your toolchain.

The ultimate winner is not the one with the strongest technology , but the one with the strongest ecosystem control.

Of course, there are also two well-known companies with a "semi-open source, semi-closed source" approach : Physical Intelligence (PI) and Genesis AI.

We will focus on them in the next part about open source.

Who is right and who is wrong among these major factions?

There is no answer in 2025 , but what is certain is that this race for basic models has only just begun.

So, after discussing so much, let's summarize the current situation in 2025: it's impressive, but implementation is still uncertain.

Musk's dream for Tesla robots is grand , but the reality is that Optimus is still struggling to climb.

On December 19th, Tesla officially released a video titled "Tesla Humanoid Robot 2025 Annual Report," which details the technological iterations and evolutionary path of the humanoid robot Optimus over the past year.

The video ends with a scene of Optimus accelerating into 2026, hinting at a larger technological leap next year.

We eagerly await the results.

Meanwhile, Figure... AI has secured a $1 billion valuation of 39 billion , but only a few dozen NVIDIA GR00T N1 robots have been released for commercial deployment. It's unclear

how many companies are actually using them.

However, we've seen impressive progress from various companies : giants like Tesla with substantial funding , startups like Figure and Dyna making rapid strides , and OpenAI and Meta quietly entering the market.

All are investing heavily in advancing the basic models of robots , leading us to believe that , especially in the home robot sector, robots helping with chores is not far off.

We believe the initial focus will be on commercial services and human-assisted tasks, as we are currently developing.

However, we don't think the home robot market is that far off.

As I mentioned earlier, you don't need a complete, universal AGI; you only need a few tasks to get the robot into the home, allowing it to perform tasks and gradually develop more capabilities through model iteration . Our goal

is to achieve this next year... (2026) We hope to have a relatively large-scale deployment in commercial scenarios. For home use, we'll consider it when the time is right . For example, with features like folding clothes, we've interviewed many friends , and everyone feels this is a very necessary function, right?

When the hardware cost drops to a level affordable for ordinary families, we might prioritize selling clothes-folding functionality to households first , and then gradually expand to other functions.

So this timeline shouldn't be too far off, probably in 1-2 years.

Would you buy a robot assistant that can fold clothes, prepare breakfast, and do cleaning for a few hundred dollars?

Okay, that's all for today.

Today we talked about the closed-source camp , but there's a group of people doing the same thing in a completely different way.

They're building an open-source model , sharing data, and believing in the power of "many grains of sand forming a tower."

They say that openness is the key to embodied intelligence.

In the next robot video, we'll discuss just how open NVIDIA's "openness" really is, and how it differs from true open source.

Why do some people say GR00T... N1 is "pseudo-open source."

Why is Physical Intelligence open source? Why would

a newly established company that just received funding release its core model for free?

What are their business models?

Open source vs. closed source: who will win?

What is the essence of this battle—is it a battle of technological routes or a battle of business models?

Don't forget to follow our account and don't miss any updates!

That's all for this video.

Your likes, comments, and shares are the best motivation for Silicon Valley 101 to create in-depth technology and business content.

I'm Chen Qian , see you in the next video!

Bye!