Startup Eon Systems demonstrates first multi-behavioral brain emulation

The US startup Eon Systems PBC has presented the first complete brain emulation in a simulated body that can display multiple behaviors. This was announced by co-founder Dr. Alex Wissner-Gross. This demonstration represents an important milestone in whole-brain emulation (WBE) research, in which a biological brain is digitally copied with neuronal accuracy and operated in simulations.

What you see here is not an animated model or an agent based on reinforcement learning. It is a real copy of a biological brain that moves a body in simulations. The mind is no longer in the machine - the machine becomes the mind.

Dr. Wissner-Gross, CoFounder Eon Systems PBC

From model to physical execution

Previous work in this field has been limited to disembodied brains or animated bodies without a neuronal basis. Eon Systems now combines for the first time a complete brain based on a biological connectome with a physically simulated body in the MuJoCo environment.

The simulation works in a closed loop: sensory inputs are processed by the brain, neural signals generate motor commands that the simulated body executes - allowing the brain emulation to show real behavior for the first time.

The demonstration video

First demonstrations of behaviors

In the current simulation, the emulated fly already shows several recognizable behaviors. These include

• Grooming: The fly cleans certain body parts, such as the antennae with the forelegs, when they are soiled or stimulated. This behavior was previously simulated in the unembodied brain model by Shiu et al. (2024).
• Feeding: Upon contact with sugar, the fly triggers a stereotypical feeding response. Sensory neurons on the legs or proboscis transmit the stimulus to the brain, which then activates the motor program for feeding. In the simulation, the fly can perceive a virtual droplet of sugar and react accordingly.
• Food forage: A more complex, targeted search for food is also simulated. The fly explores an arena, recognizes sources of smell or taste and navigates specifically towards them.
• Escape from visual threats: Stimuli such as a suddenly expanding object trigger neuronal activation in the model that would trigger escape behavior. However, this behavior is not yet implemented in the current embodied simulation.

The demonstration of these behaviors shows that the brain model in combination with a simulated body can realize initial sensorimotor loops that go beyond simple stimulus-response chains.

Limitations of current emulation

Despite the success, there are important limitations:

• Neural simplifications: The Shiu et al. model uses simplified "leaky integrate-and-fire" neurons and abstracted synapses. Many biophysical details, dendritic nonlinearities, neuronal plasticity, learning processes or hormonal states are missing. The model's responses are therefore not fully comparable to biological flies, which modulate sensory responses depending on hunger, arousal or previous experience.
• Challenges of brain-body coupling: It is still unclear how activity patterns of certain neurons are optimally translated into movements. Currently, torques, joint movements or body postures are still mapped in a very simplified way, sometimes manually assigned.
• Limited neuron interface: Only some descending neurons (e.g. DNa01, DNa02, aDN1, oDN1, giant fiber) are controlled, the entire hierarchy and redundancy of the natural system is not yet mapped.
• Research-limited approach: The current system is more of a research and demonstration platform. A complete replication of the natural behavioral repertoire requires additional learning mechanisms, more detailed motor interfaces and extensive functional data.

The embodied fly thus marks an important step for research, but is still far from being a full representation of biological reality. Above all, it serves as a platform to further explore structure-to-behavior approaches and to investigate the interaction of brain and body in simulations.

Building on previous research

The current demonstration builds on previous work:

• The FlyWire connectome of the fruit fly (Drosophila melanogaster) with over 125,000 neurons and 50 million synapses was published in Nature back in 2024 and showed motor behavior with 95% accuracy, but without a body.
• The NeuroMechFly v2 framework enabled the physical simulation of the body.
• Research into centralized brain networks helped to realistically map the coordination of body parts.

As a result, the fly's brain can now perform several natural behaviors autonomously, controlled by its own neural circuits.

Global significance and next steps

Eon Systems is pursuing the goal of developing the largest connectome-based brain emulation. In the long term, the aim is to digitally reproduce mouse and later human brains. A mouse brain contains around 70 million neurons, around 560 times as many as a fly. To this end, the team is collecting extensive connectomic data and functional recordings using calcium and voltage mapping to record the activity patterns in living tissue.

The results mark a qualitative advance in neuroscience and AI research and could lay the foundation for digital brain models at human level in the long term.