Absence of Embodied Intelligence Scaling Law: When will the "Moore's Law" in the robotics world be born?

On August 9th, on the podium of the World Robot Conference, Wang Xingxing, the founder of Unitree Technology, talked about the RL Scaling Law issue in the current field of robot motion control. He believes that when current robots learn a new skill, they often need to conduct research and teaching from scratch.

In the future, what is more hoped for is that they can continuously learn on the existing basis, making their learning speed faster and the effect better.

Currently, no one in the entire embodied intelligence industry has done a good job in the Scaling Law of reinforcement learning. The research on computing power growth and data accumulation cannot directly make robotic arms more dexterous or make bipedal robots walk more steadily.

In fact, the problem does not lie in the hardware and data we provide not being good enough, but in its “brain evolution” and continuous adaptation to the environment. Then, in the face of the lack of the Scaling Law in embodied intelligence, should we, like teaching infants, let robots keep trying and have the real ability to continuously evolve, so as to promote the birth of the “Moore's Law” in the robot world?

Computing Power Growth ≠ Intelligence Evolution: The Dilemma of Physical Bottlenecks in Embodied Intelligence

In the current field of artificial intelligence, we gradually believe that when we provide more data and computing power, we can support robots to have smarter capabilities.

For example, the development of ChatGPT seems to confirm this point - give it a larger model and more training texts, and it can write more fluent articles and answer more complex questions.

We know that in the field of virtual AI, the data is single and the rules are clear, so increasing computing power and data can improve the performance of virtual AI.

But when such a method is applied to robots, when they interact with the real world, we find that the real effect is often affected by the randomness caused by the friction of different materials, air resistance, object deformation, etc. For example, they may be tripped by some obstacles during operation.

The “mechanical ant” experiment at Harvard University vividly demonstrates this dilemma. The researchers increased the computing power of this small robot by 10 times, expecting it to better adapt to complex terrains.

But the result was disappointing: its environmental adaptability only increased by a negligible 2%. The problem does not lie in the weak chip, but in the fact that the mechanical legs of the ant cannot sense and adapt to ground changes as sensitively as real insects.

Therefore, countless physical characteristics in the real world cause countless “accidental situations” for robots during operation.

Moreover, when current robots learn new skills and adapt to new environments, their chips always require relatively large energy consumption.

This is not a problem that can be solved by continuously evolving technology to reduce energy consumption, but a fundamental defect that should be faced: robots are still using “brute-force computing” to confront the laws of physics.

This dilemma reveals a profound fact: in the field of robots, simply increasing computing power is like installing a powerful engine in a car, but forgetting that the road we are going to travel on is rough and potholed. Can it respond to the road conditions in real time?

What we hope more is that it can continuously evolve on the existing basis, just like the human biological system, to face and solve the dilemmas restricted by the physical world.

Darwinian Evolution: The Body Adaptation Wisdom of Embodied Intelligence

The key to true intelligence does not lie in whether its “brain” has learned enough. For example, in nature, an octopus does not have a centralized brain like vertebrates. Its 500 million neurons are distributed in its tentacles, enabling its eight arms to independently grasp, detect, and even “think”.

However, the behavior demonstrated by such a creature is different from the “brain” intelligence model of robots that relies centrally on data drive. Instead, it uses the body itself as part of the computation.

In the past, the progress of robots and AI mainly relied on piling up computing power and data, just like continuously providing a large amount of knowledge to a computer, hoping it would become smarter.

But in fact, the development of real intelligence, whether it is a human baby learning to walk or an animal's precise hunting, depends on the real-time interaction between our body and the environment.

Therefore, we need it to have its own adaptability in the body - just like human muscles will actively produce certain behaviors in case of emergencies.

However, robots need to adapt to different states and behaviors according to instructions and environmental changes without prior presetting.

As Wang Xingxing mentioned, he hopes that when a robot is in a new environment, it can actively explore some things to meet the instructions we provide, rather than us providing it with some specific data and it just having to follow them.

The future breakthrough may lie in the co - evolution of its “body” and “brain” like natural organisms. The algorithm no longer needs to be continuously trained for every subtle environmental condition, but can naturally adjust in the interaction.

Although this path is difficult, the evolutionary history of nature has proven its feasibility. From single - cell organisms to humans, biological evolution may provide us with a better direction.

Future robots may need to learn from organisms and develop “body intelligence” to coexist harmoniously with the physical world.

The real breakthrough is not to build more precise robotic arms, but to create more vibrant robots.

Reshape the Measurement Standard: Embodied Intelligence Needs to Embrace “Survivability”

For embodied intelligence, the real test has never been to repeatedly complete a certain fixed task, but to maintain stability in changes.

Current robots lack real anti - interference wisdom - not the ability to avoid mistakes, but the ability to quickly correct mistakes after making them.

We always laugh at the somewhat clumsy behaviors of current robots and even call their actions “old - lady behaviors”. But for such embodied intelligence that still needs continuous evolution, “fault tolerance rate” should become a new scale to measure intelligence. After all, in the real world, stability is more important than precision.

Like in childhood, when we pick up a thermos cup, we don't need to relearn the “grasping” action - we know it is a cylindrical container like a mug, and we only need to fine - tune the strength to pick it up.

But today's robots, after perfectly mastering the mug, need to start training from scratch when facing a thermos cup. The problem does not lie in their lack of “intelligence”, but in their lack of a real understanding of the concept of “cup”.

This lack of “generalization ability” exposes the core defect of current robot learning: they can only remember the grasping method for a specific object through massive data, but cannot abstract general rules.

The real breakthrough may require robots to, like children, first understand the essence of “grasping” (the relationship between shape, weight, and friction), rather than rote - memorizing the data of each object.

When the evaluation standard shifts from “precision” to “vitality”, embodied intelligence can break through the current ceiling and move towards real practical application.

References:

1. Hangzhou, the Capital of AI — Wang Xingxing's “Explosive” Speech: The ChatGPT Moment of Robots Is Coming Soon | The Transcript of Wang Xingxing's Speech at the 2025 World Robot Conference (Text + Video)

2. Machine Intelligence — Robots Are Moving Towards the ChatGPT Moment! A Tsinghua Team Discovers the Scaling Laws of Embodied Intelligence for the First Time

This article is from the WeChat public account “Rixi Vision”. Author: Xian Xian, Editor: Xian Xian. Republished by 36Kr with permission.