At the 2026 World Robot Conference (WRC), held from August 19-23 at the Beijing Yichuang International Convention and Exhibition Center, the special session "AI Large Models Empowering Robots and Embodied Intelligence: A New Industry Paradigm" took place alongside the main event. Qian Linrui, Chairman and CEO of Zhiyue Spatial Intelligence, delivered a keynote address at the forum. The following is the full transcript of his remarks.

Over the past decade-plus, we have witnessed one astonishing breakthrough after another in artificial intelligence: image recognition, language interaction, and the generation of code and video. Today, we can converse with an AI about nearly any topic imaginable. Yet one question has always lingered before us. What happens when AI steps away from the screen and truly enters the physical world?

The real world has space, time, weight, latency, and countless unpredictable changes. A cup slipping off a table edge does not wait for the model to finish its next inference. A robot losing its balance has no chance to "regenerate an answer." So today, AI is crossing a profoundly important boundary—from understanding the digital world to entering the physical world. This is what we call Physical AI: enabling intelligence to genuinely enter the world, perceive it, and act upon it.

When intelligence enters the physical world, the first thing that changes is data. In the past, we talked about data as Text, Image, and Video. But for an intelligent agent truly living in the world, these are far from sufficient. Physical AI data must simultaneously possess spatial, temporal, semantic, and physical information. More importantly, this data does not remain static. Every action the agent takes changes the world, and each change becomes new data for the next moment. Therefore, the largest training set of the future is the real-time, ever-changing real world we inhabit. How to construct such a real-time dataset is the first question I raise today.

If data changes, must models change as well? Today's large models are extraordinarily powerful. But the real world imposes entirely different demands on intelligence. It requires low power consumption, because a robot cannot carry a data center on its back. It demands continuous learning, because the world changes every day. It must process continuous time, because motion, contact, and balance all occur within milliseconds. It must also genuinely engage in physical interaction. So, does current AI have the capability to address these challenges? At the U.S. National Science Foundation conference in June 2026, hundreds of scientists agreed that current AI frameworks are hitting a bottleneck as they move into the new physical world.

How do we solve this? We know that even the most advanced organism—the human brain—maintains an extremely low power consumption of about 20 watts over the long term, continuously sensing, making rapid decisions, and functioning stably in a noisy, uncertain world. So we began to ask: if we could reconstruct a biological brain and develop algorithms consistent with the brain's computational methods, could we create a new form of intelligent computing that does not rely on massive data and massive computational power?

At the same time, the bodies that carry intelligence are also changing. Future intelligence will reside in various carriers. It can live in the digital world—becoming digital organisms, digital twins, and virtual agents. It can also empower the real world—becoming robotic arms, quadruped robots, humanoid robots, drones, and unmanned vessels. It can even reconnect with life sciences—helping us study digital brains, model organisms, brain disorders, and brain-computer interfaces. So, the intelligence of the future will not belong to just one type of body. Whether intelligence can enter different bodies, understand them, adapt to them, and continue learning is also a critical issue that needs to be solved.

So, when data changes, models change, and bodies change—Intelligence Must Change. A new Physical AI technology stack is taking shape. The AI technology stack of the past decade was very clear: Internet Data → Neural Network → Digital Application. Internet data flows into neural networks, ultimately producing applications in the digital world. Physical AI, however, is forming a different stack: Environment → Brain → Agent. When the three core elements—data, models, and agents—all change simultaneously, it also means we need new infrastructure to drive progress. Just as PyTorch once enabled developers to easily define, train, and run neural networks, Physical AI likewise requires a set of tools to build worlds, run brains, and place intelligence into different bodies.

Over the past few years, this is exactly what we have been working on. Today, I am very pleased to present it to you in its entirety for the first time. Its name is DeepSoma, known in Chinese as Zhiyue Lingwu. Through DeepSoma, we aim to accomplish three things: make the world computable, make the brain runnable, and make intelligence transferable.

Next, I want to show you directly what we have actually built. What you just saw is DeepSoma, officially released today. Let me first highlight a very important detail in the video. The fruit fly flight environment, brain simulation, and the constantly changing neuronal firing states you saw all come from DeepSoma's real engineering outputs. These visuals were not fabricated by generative AI to fake simulation results. The environment was modeled and physically simulated with high precision by our engineers using DeepSoma. The fruit fly brain was constructed as a detailed neural model based on real biological data. The continuously changing firing states on the screen are the actual neuronal discharge states recorded by our DeepSoma solver during the computation of the digital brain.

The entire DeepSoma system can be condensed into three actions: Build Worlds, Run Brains, Embody Intelligence. Within a single system, the environment, brain, and body can already operate together in a closed loop. The first step is Build Worlds. Much of this is familiar to many of you—from digital twins to the differentiable physics engines many companies are now working on, everyone is striving in the Build Worlds domain. DeepSoma's core work here is R2S plus physical constraints. Through multimodal data input, we recover geometry, semantics, pose, and contact relationships in real time, then further estimate material properties, friction, dynamics, and uncertainty, ultimately forming 4D world data composed of Space × Time × Semantics × Physics. Another key breakthrough our team achieved is in data structure—how to effectively store such multi-dimensional, multi-type data through a well-designed database. Ultimately, this part forms a closed loop through Perception → Reconstruction → Physics Embedding → Simulation → Update: Sense → Reconstruct → Embed Physics → Simulate → Update. What we want to achieve is to make the real world a truly callable, repeatable experimental data infrastructure.

The second step is what we have devoted the most time to: Run Brains. Look at the right side of the PPT—this is a real neuron under a microscope. A neuron includes dendrites, a soma, an axon, and numerous branches; it is far more complex than a simple node in an artificial neural network. Increasingly, brain science research is telling us that simplified neurons lose a great deal of storage-computation information and higher-order intelligence. So, how can such a real "biological structure" under the microscope be transformed into a model that a computer can run, thereby generating higher-order intelligence? DeepSoma encapsulates our self-developed fine-grained neuron modeling tools and neural dynamics solvers, allowing users to quickly model single neurons and progressively scale up to neural circuits, connectomes, and whole-brain dynamics. This also brings us to the question we are truly interested in: how much of intelligence comes from data, and how much from structure itself? The same knowledge, absorbed by different people, yields completely different results. The same data, entering different brain connectivity structures, produces entirely different outcomes. Therefore, DeepSoma aims to enable more researchers to actually run, compare, and explore these structures through fine-grained simulation of biological brains, driving intelligent models from data-driven approaches further toward structure-and-dynamics-driven approaches.

The third step is Embody Intelligence. DeepSoma provides a unified agent interface. Model outputs can be mapped to robotic arm actions, drone flight controls, or the behavior of a digital agent. Through DeepSoma, users can quickly perform model transfer, action decoding, and task deployment, re-establishing the relationship between a Brain and a new Body. Looking back, DeepSoma essentially converts the real world into 4D data, transforms models into dynamical systems that run continuously over time, and returns intelligence to the world through bodies. The agent acts, the world changes, new data is generated, and the brain computes again. Thus, this loop truly closes. We believe the most valuable aspects of intelligence to explore are happening within this closed loop.

Why does Zhiyue want to build such a seemingly massive system? The reason is actually simple. Because from day one, we have been studying the same question: as the old saying goes, humans are smart because they use tools. So, how exactly does intelligence gradually emerge through interaction with the real world? Today's release of DeepSoma is just a beginning for us. We hope it becomes a truly open infrastructure. If you are a neuroscientist, you can validate neural circuits and study whole-brain dynamics here. If you are an AI researcher, you can explore new foundational models beyond Transformers. If you are a robotics developer, you can put a Brain into different bodies and continuously test it in the real world. If you come from the life sciences, this can also serve as an experimental space for digital life, brain disease research, and BCI. We hope that in the future, developers will not need to rebuild all this complex infrastructure from scratch. You only need to focus on one question: what kind of intelligence do you want to create?

Finally, I want to explain why we named this platform DeepSoma. Soma is a very ancient word. It originates from ancient Greek, meaning Body. In today's neuroscience, Soma has a more specific meaning—the cell body of a neuron. This precisely corresponds to the two scales at which we understand intelligence. At the micro level, it is a neuron, a dendrite, and the complex connections formed among hundreds of millions of neurons. At the macro level, it is the body, the environment, and how a living being perceives, learns, and acts in the real world. So we added the word "Deep" before Soma. DeepSoma represents our aspiration to start from the most authentic, deepest structures of life and understand intelligence layer by layer—but "Deep" also carries another meaning. We do not know where the boundaries of intelligence lie. Today, we replicate a neuron and run a fruit fly's brain; in the future, Zhiyue will gradually replicate more complex brains and explore forms of intelligence we cannot yet imagine. So DeepSoma is also an invitation. We want to hand this toolset to researchers and developers around the world to jointly explore the deeper frontiers of intelligence. From understanding a single neuron to understanding a whole brain, from understanding life—to creating new intelligence. Zhiyue Spatial Intelligence. DeepSoma, Zhiyue Lingwu. Welcome to join us in entering the new paradigm of Physical AI and exploring more possibilities. Thank you.