Tsinghua AIR's Cao Ting: The Biggest Blind Spot in Embodied Intelligence Is the Physical Agent

Introduction: A Sobering Perspective Amid the Embodied Intelligence Wave

In 2025, the embodied intelligence sector continues to surge, with capital and technological resources pouring into everything from humanoid robots to autonomous driving. Yet amid the frenzy, Cao Ting, a professor at Tsinghua University's Institute for AI Industry Research (AIR), has raised a thought-provoking point — the most overlooked issue in embodied intelligence is the "physical agent" itself.

In a recent exclusive conversation, Professor Cao systematically laid out her in-depth thinking on the development path of embodied intelligence and, for the first time, revealed that her team is preparing to incubate a new company aimed squarely at "continuously evolving physical agents."

Core Thesis: The Industry Is Overly Focused on Algorithms While Ignoring the Essence of the Physical World

In the conversation, Professor Cao was blunt: a significant cognitive bias exists in the current embodied intelligence field. Vast research resources are concentrated on large models, perception algorithms, and planning-decision layers, while insufficient attention is paid to the agent's genuine ability to interact in the physical world.

"We've spent too much effort teaching robots to 'think' and forgotten to let them truly 'live' in the physical world," said Cao. She argued that a physical agent is not merely a hardware carrier but should be a complete system capable of perceiving environmental changes, adapting to unknown scenarios, and continuously evolving through sustained interaction.

She further explained that many current embodied intelligence systems perform impressively in laboratory settings but suffer dramatic performance drops once deployed in real, unstructured environments. The root cause is that these systems lack deep understanding of the physical world's complexity and the ability to self-iterate in real environments.

"A true physical agent should be like a living organism — able to accumulate experience, correct behavior, and even reconstruct the boundaries of its own capabilities through continuous interaction with the environment," Cao emphasized. This capacity for "continuous evolution" is the critical bottleneck for embodied intelligence to achieve large-scale real-world deployment.

Deep Dive: Why Are Physical Agents So Important?

From a technical standpoint, the core challenges of physical agents span at least three dimensions:

First, the uncertainty of physical interaction. Unlike the digital world, the physical world is filled with unpredictable factors such as variations in friction, material differences, and lighting fluctuations. Traditional end-to-end learning methods struggle to fully cover these long-tail scenarios. Agents need the ability to learn online and adapt in real time.

Second, the co-evolution of body and intelligence. Cao pointed out that current research often treats the "body" and the "brain" as separate — hardware teams work on hardware, algorithm teams work on algorithms. But in nature, an organism's body structure and intelligence evolve in tandem. Physical agent design should follow the same principle, creating a unified evolutionary loop across morphology, perception, and decision-making.

Third, balancing continuous learning with safety. An agent capable of continuous evolution inevitably faces the challenge of ensuring behavioral safety while acquiring new capabilities. This is not only a technical challenge but also a matter of system reliability engineering in practice.

From an industry perspective, the maturity of physical agents directly determines the commercialization ceiling of embodied intelligence. Whether for home service robots, industrial collaborative robots, or specialized task robots, the ultimate competition is not about peak performance on a single task but about long-term stable operation and autonomous adaptability in complex real-world environments.

New Company in Preparation: From Academic Insight to Industrial Practice

Notably, Professor Cao and her team have not stopped at academic discourse — they are putting this core philosophy into industrial practice. The team is currently planning to incubate a new company focused on "physical agents capable of continuous evolution."

While the company is still in the preparatory stage and specific product forms and business models have not been disclosed, based on Professor Cao's research trajectory, the company will likely focus on the core system architecture of physical agents — including an integrated perception-decision-execution framework with online adaptation capabilities and a foundational platform to support continuous learning and evolution of agents in real-world environments.

As one of China's top intelligent industry research institutions, Tsinghua AIR has deep expertise in embodied intelligence, autonomous driving, and AI systems. The incubation of a startup from AIR also reflects an accelerating trend of translating academic achievements into industrial applications.

Outlook: Physical Agents May Become the Next Major Battleground in Embodied Intelligence

Professor Cao's perspective offers the industry an important new direction for consideration. As the capability boundaries of large models continue to expand and algorithmic gaps gradually narrow, the competitive focus of embodied intelligence may shift from "whose model is stronger" to "whose agent can better adapt to the real world."

The concept of the physical agent is essentially a call for a more systematic and physics-respecting development paradigm for embodied intelligence. It demands that researchers and entrepreneurs focus not only on algorithmic precision but also on the agent's survival capability and evolutionary potential as a complete physical entity.

As Professor Cao's team gradually brings its new company to fruition, the physical agent direction is poised to attract greater attention and resource investment. At this critical juncture where embodied intelligence moves from the laboratory to the real world, this may be exactly the dose of clear-headed thinking the industry needs most.