Reconstructing Real-World Scenarios into a Continuously Updated, Computable 4D Digital World
By Yuzhong, from Aofeisi
QbitAI | WeChat Official Account: QbitAI
Just as a digital fruit fly sent “uploading the brain” trending online, a Chinese team has set its sights on the three hardest pieces to complete:
They are not satisfied with LIF simplified neurons, not satisfied with preset physics-simulation environments, and not satisfied with creating merely a digital animal that can move. They aim to connect detailed brain models, a digital world generated from real-world environments, and interfaces that work across bodies.
In March this year, Eon Systems unveiled a “digital fruit fly” that forages and grooms its antennae in a virtual environment. Based on the connectome of an adult fruit fly, it involves approximately 140,000 neurons and 50 million synaptic connections, combined with a biomechanical body featuring 87 independent joints. For the first time, neural activity, body movement, and environmental feedback formed a complete loop.
Once the video was released, the claim that “the fruit fly’s brain has been uploaded” quickly went viral. Eon Systems itself, however, has remained fairly measured: The current system uses a simplified leaky integrate-and-fire neuron model, some mappings between the brain and body are still manually specified, and only a limited range of sensory inputs and behaviors is covered. It is better understood as a highly compelling research platform and proof of concept than as a perfectly replicated biological fruit fly.
It was at this juncture that Chinese embodied-intelligence company ZhiYue Spatial Intelligence unveiled DeepSoma (ZhiYue Lingwu). Rather than defining the product as “another digital animal,” the company is attempting to build a more fundamental whole-brain simulation platform: turning the real world into a computable environment, translating biological constraints and whole-brain connectivity structures into executable models, and then connecting those models to different bodies.
In a nutshell: Eon Systems demonstrated “how a simplified fruit fly brain can drive a physics-simulated virtual body.”
DeepSoma presents a more ambitious result: placing a fully reconstructed fruit fly brain directly into the real world and a real body.
So, Is It Really an “Enhanced Version” of Eon Systems?
A simplified digital fruit fly is already exciting enough. For DeepSoma, however, it is not the endpoint—it is targeting the three hardest pieces to complete.
- Model layer:
Eon’s publicly disclosed approach is an LIF fruit fly brain constrained by its connectome; DeepSoma, by contrast, emphasizes detailed biophysical modeling of neurons.
Eon uses the leaky integrate-and-fire (LIF) model to represent approximately 140,000 neurons and 50 million synaptic connections, creating a brain–body loop in the fruit fly with relatively simplified and efficient neural dynamics.
DeepSoma models further down, inside individual neurons: biophysical mechanisms including dendrites, somas, membrane potentials, ion channels, and synapses are all converted into computable and observable states, which are then connected through the connectome to form neural circuits and a whole-brain network.
The former gets the fruit fly brain’s circuitry running; the latter attempts to enter every computational unit within that circuitry and determine how intelligence is actually computed.
- World layer:
Eon’s fruit fly completes a perception–action loop in a physics-simulation environment; DeepSoma, by contrast, first reconstructs real-world scenes into a continuously updated, computable 4D digital world.
Eon’s public demonstration relies on the MuJoCo physics engine and a virtual fruit fly body, which receives sensory inputs, generates actions, and obtains feedback in a preset simulation environment. DeepSoma is not satisfied with simulation alone. Instead, it places greater emphasis on reconstructing geometric, semantic, spatiotemporal, and physical information from the real world to provide the brain with a more realistic training ground.
- Platform layer:
Eon presents a highly compelling digital fruit fly case study; DeepSoma, by contrast, integrates real-world modeling, detailed whole-brain computation, and cross-embodiment Agent integration into a complete platform.
DeepSoma summarizes this pipeline as Build Worlds, Run Brains, Embody Intelligence:
The world layer provides an interactive environment, the brain layer runs dynamics ranging from detailed neurons to a whole-brain connectome, and the Agent layer maps the same brain model onto digital animals, biological experimental systems, robotic arms, or humanoid robots. The body’s actions change the world, the world’s new state then enters the brain again, ultimately closing the loop: Environment → Brain → Agent → Environment.
Of course, broader platform boundaries and more detailed neuron models also mean a higher bar for validation. Can the platform reliably run larger whole-brain networks? How closely do the simulation results correspond to real neural activity? Can brain–body mapping reduce the need for manually defined rules? Can cross-body transfer be reproduced on public tasks? These are all questions that ZhiYue Spatial Intelligence, a young company, will need to explore step by step and answer with greater precision.
Physical AI May Really Need Its Own “PyTorch Moment”
Today, Physical AI is emerging as a new direction for the AI industry.
If intelligence is truly to enter the physical world, it needs not a larger model or more data, but an entirely new AI technology paradigm!
ZhiYue Spatial Intelligence has entrusted DeepSoma with the task of creating this new technological paradigm.
DeepSoma’s positioning can be summed up as: Build Worlds. Run Brains. Embody Intelligence—making the world computable, the brain executable, and intelligence transferable.
This naturally brings PyTorch to mind. PyTorch is one of the most widely used development frameworks for deep learning today. It does not dictate what a neural network must look like; instead, it provides a common foundation for defining, training, and running models. DeepSoma aims to do something similar, except that its computational objects extend beyond networks in the digital world to a closed loop composed of physical environments, biological brain structures, and intelligent agents.
Conclusion
As large models become increasingly capable of speaking, seeing, and reasoning, the next truly difficult challenge is no longer making them answer more elegantly on a screen. It is enabling them to work continuously and reliably in a constantly changing world. This may seem like one of the most basic survival abilities in biology, but it poses a major test for AI.
At this moment, more and more teams are exploring the intelligence secrets of Physical AI through biological brains. Eon Systems brought “brain uploading” into the public eye with a digital fruit fly; DeepSoma is betting that what will truly matter in the future is not uploading a particular brain, but building a computational system that enables the world, the brain, and the body to evolve together.
If this path succeeds, the next leap in Physical AI may come not only from larger models and more data, but also from an older and more efficient computational prototype—life itself.