At the "Physical AI and Embodied Intelligence Robot Innovation Ecosystem Exchange" event, held as a sideline forum of the China International Fair for Trade in Services, Zhang Zhan, Co-founder and COO of Zhiyue Spatial Intelligence, presented a keynote address.
Zhang Zhan, representing Zhiyue Spatial Intelligence, outlined his team's decade-long focus on the intersection of brain-inspired intelligence and embodied intelligence. Their primary goal is to break free from the data bottleneck of traditional AI, re-engineer the foundational logic of AI based on the computational mechanisms of biological neurons, and develop a next-generation intelligent paradigm with independent Chinese intellectual property.
Currently, the industry's dominant approach for non-embodied tasks like language and image recognition relies on scaling laws — amassing data and computational power to spur intelligence. However, when this strategy is applied to embodied intelligence, where robots interact with the physical, three-dimensional world, it hits a fundamental barrier. Mainstream AI is a passive intelligence, heavily dependent on pre-training, and rife with limitations: poor environmental awareness, insufficient spatial reasoning, weak cross-scenario generalization, unnatural human-robot interaction, and poor flexibility. It consistently fails to breach the boundaries of general intelligence.
In contrast, the human brain operates on just 20 watts of power, a minimal amount of energy that drives 860 billion neurons. It can perceive new environments, adapt to new scenarios, and dynamically capture changes without relying on massive datasets or pre-training. Ilya, the former CTO of OpenAI, has shared similar views, suggesting the ultimate form of general intelligence is an agent with biological learning mechanisms capable of autonomous learning and evolution in its environment.
This leads to a core conclusion: true general intelligence is not trained but emerges naturally from biological structure. The source of intelligence is the combined function of a single neuron's computational logic, the network's connection structure, and the body morphology evolved over time. Following this, the team is advancing on two frontiers: network scale, from a simple roundworm's 300 neurons up to a mouse's 100 million and ultimately the human brain's 86 billion; and the complexity of individual neurons, enhancing the computational fidelity of dendrites and axons.
Zhiyue Spatial Intelligence positions itself as the world's only company dedicated to fine-grained modeling of complete biological brains. Its mission is to integrate this with embodied intelligence, providing "large and small brain systems" for robots and advancing AI from a parameter-driven to a structure-driven stage.
At the end of 2024, the company achieved its first milestone by fully replicating the roundworm's entire 302-neuron network using precise mathematical equations. This work was published on the cover of a Nature sub-journal. Crucially, they were the first to complete the full loop in academia: from environmental perception, through neural processing, to driving every muscle for action and feeding back into the environment, validating the complete path of biological neural structure driving behavior.
Building on this, the team has since upgraded to the more advanced fruit fly nervous system. While the EonSystem company, recently gaining viral attention and praise from Elon Musk, only achieved a basic digital network simulation, Zhiyue Spatial Intelligence has precisely replicated fruit fly neurons and their entire connectome, creating more accurate and dynamic biological neuron computational models. This represents a full surpass of current industry-leading technology. Using their self-developed simulator and the brain-inspired training platform DeepSoma, launched at the World Robot Conference, they embedded the digital fruit fly into a dynamic physical environment. The fly, without any data pre-training or manual rule programming, autonomously perceives its environment, avoids obstacles, and adapts its movement, perfectly replicating real biological intelligence.
The most significant breakthrough is applying this original biological system to actual robots. By transplanting the mature neural structures of the roundworm and fruit fly onto industrial robots, the reliance on massive data training and heavy computation is eliminated. The robot can autonomously emerge new adaptive behaviors in real work scenarios, such as reducing jitter within centimeter or even millimeter precision for precise movements. This establishes an industry-unique technical paradigm where behavior emerges from structure.
This technology has been validated in two typical scenarios: semiconductor SMT line-side warehouse pallet pick-and-place operations, and retail order picking. The team is now building more complex neural circuits to form an "atomic action library," aiming to drive dual-arm or multi-collaborative-arm tasks with low power consumption and without pre-training.
In a global comparison, the team's advantages are clear. MIT-backed Liquid AI extracts a structure of just 17 neurons from the roundworm to enable a drone to fly steadily, avoid obstacles, and track dynamic targets without pre-training. EonSystem, after replicating the fruit fly's neural structure, can drive agents to forage, avoid obstacles, and groom themselves in simulations. Zhiyue Spatial Intelligence leads in three dimensions: the fidelity of individual neurons, the completeness of the connectome replication, and the successful deployment in real physical environments.
The company's progress is supported by an exceptional team and strong academic-industry backing. The founder, Dr. Qian, holds dual doctorates from Columbia University in Applied Mathematics and Applied Psychology, previously worked on machine vision at Boston Dynamics, and has extensive experience in mass-market product development. Zhang Zhan, the speaker, is a graduate of the University of Chinese Academy of Sciences, a serial entrepreneur who has led multiple successful funding rounds. The CTO, Sun, has a background in basic physics from Peking University and a mechanical engineering PhD from Columbia University; he is a distinguished talent in Beijing and was previously the algorithm director for Xiaomi's first-generation CyberDog and CyberOne robots.
The project has received significant recognition, including a second prize at the 2024 HiCool Global Innovation and Entrepreneurship Competition, a top-eight finish in the national finals of CCTV's "Win in AI+", and inclusion in the national reserve pool for disruptive innovative technologies. Zhiyue Spatial Intelligence also actively contributes to national policy, participating in the Ministry of Industry and Information Technology's brain-inspired intelligence industrial map and future technology initiatives. The commercial sector has shown strong confidence, with the company securing funding across six rounds within two years, totaling hundreds of millions in RMB.
In conclusion, the vision of Zhiyue Spatial Intelligence is clear: to use China's original brain-inspired technology to redefine the global general intelligence paradigm. The team firmly believes that moving from data-driven to structure-driven models is the inevitable path to the next generation of general intelligence, and they are committed to creating low-power, highly adaptive intelligent entities that can be used widely, carving an independent and controllable new path for China in the global intelligence race.
Comments