Closed Loop Neuromechanical Simulation of Embodied Agents

This direction is merged into another direction: embodied neural simulation. Its last version stays readable and citable here.

Version 1 of 126 papers34 conceptsfrozen 2026-10-04

In traditional computational neuroscience, researchers often treat the brain as a "processor in a vat"—an isolated system that receives abstract inputs and produces mathematical outputs. However, biological intelligence did not evolve in a vacuum; it evolved to move a body through a physical world. A growing body of research argues that to truly understand neural circuits, we must simulate them within the context of embodiment: a closed-loop system where a brain model controls a physical or virtual body that, in turn, provides sensory feedback from its environment.

The Problem of the Disembodied Brain

Most neural models operate in an "open-loop" fashion. They might accurately simulate how a neuron fires in response to a stimulus, but they lack the sensory-motor feedback loops that define real-world behavior. As Tatyana O. Sharpee, Frances K. Skinner, and colleagues noted in 25th Annual Computational Neuroscience Meeting: CNS-2016, there is a critical need to transition toward embodied agents to evaluate how neural heterogeneity and complexity actually drive functional performance.

Without a body, we cannot see the "shape" of neural degradation. For instance, James Kunert and J. Nathan Kutz demonstrated in Functionality and Robustness of Injured Connectomic Dynamics in C. elegans that decoupling neural dynamics from physical constraints makes it impossible to predict how microscopic axonal damage manifests as macroscopic behavioral deficits.

Robots and Virtual Bodies as Scientific Tools

To bridge this gap, researchers like Auke Ijspeert and Sten Grillner use bioinspired robots as physical "scientific instruments." In Foreword for the special issue on Lamprey and Salamander Robots and the Central Nervous System, they explain that embodying biological control models in hardware allows researchers to observe emergent properties—like the transition from swimming to walking—that result from the interplay between neural circuits and real-world physics like hydrodynamics and friction.

When physical robots are too cumbersome, high-fidelity virtual "digital twins" take their place. The Neurorobotics Platform (NRP), developed by Alois Knoll, Egidio Falotico, and others, provides a web-based environment where spiking neural networks (SNNs) can control musculoskeletal models. This allows for "Closed-Loop Neuroscience," a paradigm where, as Florian Walter argues in Closed-Loop Neuroscience - A Vision, neural activity patterns emerge naturally from the perception-cognition-action cycle.

The Neuromechanical Translation Layer

The core technical challenge of this research is the translation of microscopic neural spikes into macroscopic physical torque. This requires a "neuromechanical" model that simulates several layers:

  1. Neural Dynamics: Spiking neurons or cellular simulations.
  2. Activation Dynamics: Converting spikes into muscle excitation (EMG).
  3. Musculoskeletal Dynamics: Calculating how muscles pull on bones to create joint torque.
  4. Environmental Physics: Accounting for external forces like gravity or friction.

Massimo Sartori and Utku Ş. Yavuz established a direct causal link in this chain by decoding motor neuron spikes in humans and feeding them into subject-specific musculoskeletal models to predict ankle torque in In Vivo Neuromechanics. Similarly, the NeuroMotion platform by Shihan Ma and Arnault H. Caillet uses deep learning to synthesize realistic surface EMG signals from these underlying fiber dynamics, providing a tool for training neural interfaces in silico.

Emergent Behavior in Closed Loops

The ultimate validation of this approach is the emergence of complex behavior without hard-coded rules. In An integrative data-driven model simulating C. elegans brain, body and environment interactions, Ning Wang introduced BAAIWorm, a model that reproduces realistic foraging. The worm doesn't have a "foraging script"; instead, the behavior emerges from the continuous feedback between its biophysical neurons and its 3D environment.

This same principle applies to learning. Yufei Wu and Aldo A. Faisal showed in Reproducing Human Motor Adaptation in Spiking Neural Simulation that by coupling a spiking controller with a physiological muscle model, they could replicate how humans adapt to physical force fields—a feat impossible for a disembodied model.

Scaling to Supercomputers

As these models grow to include millions of neurons and complex musculoskeletal systems, the computational demand skyrockets. Benedikt Feldotto and colleagues have demonstrated that the Neurorobotics Platform can scale on High-Performance Computing (HPC) infrastructure, as detailed in Deploying and Optimizing Embodied Simulations of Large-Scale Spiking Neural Networks on HPC Infrastructure.

In a landmark experiment, Yusuke Kuniyoshi and Carlos Enrique Gutierrez connected a massive spiking brain model to a mouse musculoskeletal body across the internet, utilizing the supercomputer Fugaku. This architecture allows researchers to leverage global computing resources to drive real-time motor control and receive sensory feedback, finally bringing the "brain in a vat" into a body capable of interacting with the world.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_closed_loop_neuromechanical_simulation_o_ec687f52,
  title = {Closed Loop Neuromechanical Simulation of Embodied Agents},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/closed-loop-neuromechanical-simulation-of-embodied-agents/rdr_ec687f525ba94d4491527c38d18f64b2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_closed_loop_neuromechanical_simulation_o_ec687f52_v1,
  title = {Closed Loop Neuromechanical Simulation of Embodied Agents},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/closed-loop-neuromechanical-simulation-of-embodied-agents/rdr_ec687f525ba94d4491527c38d18f64b2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}