Closed-Loop Neuromechanical Simulation

32 concepts24 papers2 proposals

For decades, computational neuroscience treated the brain as an isolated information processor. Models simulated complex spiking neural networks (SNNs) or central pattern generators (CPGs) in an open-loop vacuum, feeding them artificial inputs and observing their outputs without any physical consequence. As Sharpee et al. note in 25th Annual Computational Neuroscience Meeting: CNS-2016 and Walter et al. argue in Closed-Loop Neuroscience - A Vision, this lack of embodiment ignores the evolutionary context of the nervous system, which designed the brain specifically to control a physical body.

When we decouple neural dynamics from physical embodiment, we implicitly assume that neural activity patterns map directly and cleanly to behavior. However, Kunert et al. demonstrate in Functionality and Robustness of Injured Connectomic Dynamics in C. elegans: Linking Behavioral Deficits to Neural Circuit Damage that this abstraction limits our ability to predict how neural distortions manifest as actual physical locomotion. A biological organism does not move in a mathematical vacuum; its actions are fundamentally governed by muscle activation dynamics, skeletal geometry, gravity, and environmental friction.

Furthermore, isolating neural models masks the critical challenge of stability. As Hellyer et al. show in From homeostasis to behavior: Balanced activity in an exploration of embodied dynamic environmental-neural interaction, introducing an interactive environment creates a continuous sensory-motor loop that can rapidly destabilize ungrounded neural models. Understanding brain function requires a closed-loop simulation where the neural network is fully embodied within a physically realistic environment, forcing neural activity to emerge from realistic sensory feedback and motor execution rather than arbitrary inputs (Feldotto et al. in Deploying and Optimizing Embodied Simulations of Large-Scale Spiking Neural Networks on HPC Infrastructure).

The Neuromechanical Loop as an Emergent Controller

To bridge the gap between neural activity and behavior, researchers utilize forward neuromechanical simulation: embedding hypothesized neural architectures into virtual body models to evaluate their ability to generate biorealistic behavior (Wang-Chen et al. in The embodied brain: Bridging the brain, body, and behavior with neuromechanical digital twins). In this paradigm, movement is not merely hard-coded in the brain; it emerges from the dynamic interaction between neural commands, muscle elasticity, and environmental feedback.

Consider animal locomotion. Ijspeert highlights in Central pattern generators for locomotion control in animals and robots: A review how coupling CPG control systems with biomechanical simulations allows us to study how neural control signals translate into physical movement. By building a unified architecture that integrates abstract neural oscillators with linear Hill-type muscle models, Ijspeert et al. demonstrated in Effects of muscle dynamics and proprioceptive feedback on the kinematics and CPG activity of salamander stepping that a salamander's transition from swimming (traveling waves) to walking (standing waves) does not require a complex, top-down neural switch. Instead, the shift in gait can emerge naturally from changes in muscle stiffness and sensory feedback loops interacting with the environment.

This closed-loop interaction does more than just execute commands; it actively shapes how the brain itself is organized. The constraints of physical movement and body dynamics are primary drivers of neural representational structure in motor and sensory regions. For instance, Wang-Chen et al. note that artificial neural networks trained via reinforcement learning to control virtual musculoskeletal bodies accurately predict the actual neural activity patterns observed in biological brains during similar tasks.

Translating Neural Codes to Physical Torque

Validating these models requires establishing a direct, causal pipeline from neural firing to physical mechanics. This can be approached from both open-loop and closed-loop perspectives:

[Cortical/CPG Activity] ──> [Spinal Cord Model] ──> [Muscle Activation (EMG)] ──> [Physics Engine / Torque] ──> [Environmental Feedback] ──┐
       ▲                                                                                                                                │
       └────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘

In a feed-forward setup, Sartori et al. demonstrated in In Vivo Neuromechanics: Decoding Causal Motor Neuron Behavior with Resulting Musculoskeletal Function that decoding individual motor neuron spike trains from living humans and feeding them into a subject-specific musculoskeletal model can accurately predict joint torque. This open-loop architecture establishes a direct causal link between neural commands and mechanical output by relying solely on initial calibration, proving that biophysical muscle models can reliably translate raw neural codes into mechanical output.

To explore more complex, closed-loop scenarios, Mascaro et al. developed an "Embodied Brain" framework in Experimental and Computational Study on Motor Control and Recovery After Stroke: Toward a Constructive Loop Between Experimental and Virtual Embodied Neuroscience. Here, real-world cortical recordings from mice are processed through a digital spinal cord model to drive a virtual musculoskeletal system. By accounting for physical constraints like muscle activation and environmental friction, this digital twin allows researchers to see exactly how neural changes—such as those occurring during stroke recovery—manifest as motor outcomes.

This physical grounding is essential for testing new neuroscientific hypotheses. For example, Strohmer et al. proposed a model of spinal motor circuits in Sparse Firing in a Hybrid Central Pattern Generator for Spinal Motor Circuits characterized by "sparse firing" rather than hyper-synchronized bursts. The functional validity of such low-frequency signals remains unproven in isolation; it is only by integrating them with musculoskeletal simulations that the authors could verify whether these sparse patterns are physically capable of coordinating complex locomotion. Similarly, Kim et al. developed the Neural Interactome: Interactive Simulation of a Neuronal System to allow researchers to alter specific neurons on the fly and immediately observe how those changes ripple through the network to produce physical movement, highlighting the necessity of integrating neural dynamics with biomechanical body models.

Scaling Up to High-Fidelity Digital Twins

Simulating both a complex brain and a realistic physical body simultaneously is computationally punishing. It requires specialized software frameworks and high-performance computing (HPC) infrastructure to close the loop in real time.

The Neurorobotics Platform (NRP), detailed by Falotico et al. in Connecting Artificial Brains to Robots in a Comprehensive Simulation Framework: The Neurorobotics Platform and developed as part of the Human Brain Project (Knoll et al. in The Human Brain Project—Synergy between neuroscience, computing, informatics, and brain-inspired technologies), provides a standardized web-based environment to connect spiking neural networks to virtual robotic or musculoskeletal bodies. Within this platform, researchers can use optimization techniques, such as genetic algorithms, to map electromyography (EMG) data to mechanical joint torques, revealing how biological systems exploit muscle redundancy (Feldotto et al. in Evaluating Muscle Synergies With EMG Data and Physics Simulation in the Neurorobotics Platform; Knoll et al. in Editorial: Neuromorphic engineering for robotics).

As these neural models scale up to millions of neurons, the computational demands require distributed architectures. Feldotto et al. demonstrated near-linear scaling of large-scale SNNs within physically realistic environments by deploying the NRP on HPC infrastructure (Deploying and Optimizing Embodied Simulations of Large-Scale Spiking Neural Networks on HPC Infrastructure). Taking this a step further, Kuniyoshi et al. successfully integrated a massive, multi-region spiking brain model with a mouse musculoskeletal body model across geographically distributed hardware in Embodied bidirectional simulation of a spiking cortico-basal ganglia-cerebellar-thalamic brain model and a mouse musculoskeletal body model distributed across computers including the supercomputer Fugaku.

By utilizing standard web protocols to link high-performance computing resources with physics engines, this architecture allows supercomputers to drive real-time motor control and receive reciprocal sensory feedback from a remote body. This realizes the vision of a complete, biorealistic digital twin: a tool that proves why neural circuits are structured the way they are by testing them against the uncompromising laws of the physical world.

Key Foundational Work

Dolphin-like Swimming Modeling for a Biomimetic Amphibious Robot (2011) — contributed 2 concept elements to this direction.
Central pattern generators for locomotion control in animals and robots: A review (2008) — contributed 1 concept elements to this direction.
Effects of muscle dynamics and proprioceptive feedback on the kinematics and CPG activity of salamander stepping (2011) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_embodied_neural_simulation_461536b7,
  title = {embodied neural simulation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/embodied-neural-simulation/rdr_461536b7cb894361b77f05b02abe4da1},
  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_embodied_neural_simulation_461536b7_v1,
  title = {embodied neural simulation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/embodied-neural-simulation/rdr_461536b7cb894361b77f05b02abe4da1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake