Technical Summary: The Embodied Brain: Bridging the brain, body, and behavior with biorealistic neuromechanical models

Problem Statement

Animal behavior arises from the continuous, closed-loop interaction between the nervous system, the body, and the environment. However, traditional experimental neuroscience faces two fundamental limitations in studying these interactions:

1. Observational Barriers: Obtaining comprehensive neuronal recordings is difficult, particularly in large animals where genetic targeting is challenging or where high-density arrays cannot resolve individual cell identities. Furthermore, measuring muscle physiology often interferes with natural behavior, and internal biomechanical variables (e.g., joint torques, internal forces) are often inaccessible or obscured by physical constraints.
2. Perturbation Limitations: Establishing causality typically requires perturbations. Existing methods are often either too local (e.g., optogenetics targeting few neurons) or too broad (e.g., pharmacological or genetic ablations of entire regions). There is a lack of experimental capability to perform patterned, high-specificity manipulations of large neuronal populations or to test counterfactual hypotheses (e.g., "what if" scenarios) without physically altering the animal.

Consequently, studying any single component (nervous system, body, or environment) in isolation provides only a partial view of the integrative whole.

Methodology

The paper reviews the state-of-the-art in biorealistic neuromechanical models, which serve as a computational framework to bridge these gaps. These models are defined by three hierarchical levels of complexity:

• Biomechanical/Body Models: Morphologically realistic representations of animals (or subparts) with articulated joints, physical geometry, mass, and contact dynamics.
• Neuromechanical Models: Biomechanical models coupled with emulated neural dynamics. These integrate biological actuators (motor neurons, muscles) and sensors (proprioceptors, tactile sensors) to form a closed-loop system.
• Neuromechanical Digital Twins: Specific instances of neuromechanical models updated continuously with real-time experimental measurements.

The review categorizes the application of these models into two primary methodological directions (illustrated in Figure 3):

1. Reverse Simulation (Inference): Experimental behavior (e.g., pose estimation) is replayed in silico using inverse kinematics and dynamics. This allows for the inference of latent variables such as internal forces, muscle activation patterns, and sensory feedback that are difficult to measure directly. This process often utilizes Imitation Learning (IL) or optimization criteria (e.g., minimizing muscle activation) to resolve the redundancy problem inherent in musculoskeletal systems.
2. Forward Simulation (Hypothesis Testing): Hypothesized neural controllers (ranging from Central Pattern Generators to connectome-constrained artificial neural networks) are embedded into body models. The resulting simulated behaviors are compared against experimental data to validate or refute specific neuroscientific hypotheses regarding circuit dynamics, timing, and synchronization.

The review highlights the integration of differentiable physics engines (e.g., MuJoCo Warp, Brax), differentiable neuron modeling (e.g., Jaxley), and GPU-accelerated simulation to achieve high-fidelity, scalable, and differentiable neuromechanical simulations.

Key Contributions and Findings

The paper synthesizes recent advances (primarily from the last five years) across five thematic areas:

1. Inference of Unmeasurable Variables

Biorealistic models enable the reconstruction of hidden physiological states. By replaying recorded behaviors, researchers have inferred:

• Internal Forces and Muscle Synergies: Identifying co-activation patterns and muscle synergies in Drosophila and rodents that are not directly observable.
• Sensory Feedback: Reconstructing proprioceptive signals (e.g., spindle responses) and visual system activities to test hypotheses about sensory encoding roles.
• Neural Control Signals: Using inverse dynamics controllers to identify candidate brain regions (e.g., striatum, motor cortex) responsible for specific control computations by comparing simulated and real neural activities.

2. Hypothesis Testing via Forward Engineering

Simulations allow for the systematic testing of neural control theories that are difficult to probe experimentally:

• Closed-Loop Behaviors: Models have successfully replicated complex sensorimotor behaviors, such as obstacle avoidance in flies and zigzag movements in C. elegans, by integrating visual, olfactory, and mechanosensory feedback.
• Robustness and Adaptation: Simulations have tested the robustness of controllers against perturbations (e.g., spinal transection in lamprey models) and explored how sensorimotor prediction errors drive adaptation in mice.
• Developmental Emergence: Models have replicated the emergence of distinct swimming patterns in developing zebrafish by modeling putative neural circuits across developmental stages.

3. Synergies with Robotics and NeuroAI

The paper emphasizes a bidirectional flow of knowledge:

• Neuroscience to AI: Biological principles (e.g., hierarchical control, decentralized feedback, intrinsic motivation) inform the design of robust artificial agents and reinforcement learning algorithms.
• AI to Neuroscience: Machine learning algorithms, particularly Imitation Learning and Reinforcement Learning, are embodied in realistic body models to test if they can explain observed animal behaviors, serving as a validation tool for neuroscientific hypotheses.
• Robotics: Neuromechanical models have inspired physical robot designs, such as hexapod robots based on Drosophila anatomy and soft robots based on jellyfish resonance.

4. Healthcare Applications

The review details the translation of these models to clinical settings:

• Pathology Modeling: Simulating pathological gaits (e.g., in stroke or cerebral palsy patients) to identify whether impairments stem from neural or biomechanical causes.
• Individualized Medicine: Creating "digital twins" constrained by patient-specific data (CT scans, EMG) to predict load distribution, optimize assistive devices, and plan rehabilitation strategies.

5. Current Limitations and Future Directions

The paper acknowledges significant challenges, including:

• Model Realism: Current body models often rely on rigid bodies and simplified sensory/muscle systems. Fluid dynamics for flight/swimming and complex soft-body interactions remain computationally costly and difficult to model.
• Parameter Tuning: Distinguishing between parameters that improve simulation realism (modeling artifacts) and those representing true biological quantities is a critical conceptual challenge.
• Data Constraints: While connectomics and high-resolution scans are improving, many models still lack sufficient anatomical and physiological constraints.

Significance and Claims

The authors posit that biorealistic neuromechanical models represent a transformative tool for neuroscience, moving the field from purely observational studies to active probing of brain-body-environment interactions.

• Bridging Levels: These models provide a unified framework to link neural circuits, biomechanics, and behavior, allowing researchers to bridge multiple levels of biological organization.
• Counterfactual Experimentation: By enabling "what if" scenarios and unrestricted perturbations (both neural and mechanical), these models allow for the testing of counterfactual hypotheses that are ethically or technically impossible in vivo.
• Accelerating Discovery: The coupling of experimental studies with active probing of neuromechanical surrogates is envisioned to significantly accelerate progress in understanding sensorimotor control.
• Foundation for NeuroAI: The paper claims these models provide a foundational framework for "NeuroAI," facilitating the development of next-generation embodied AI systems that function robustly in real-world environments.

Ultimately, the paper argues that while challenges in realism and computational cost remain, the integration of high-fidelity simulation with experimental data will shift neuroscience from post-hoc analysis to a more dynamic, hypothesis-driven science of embodied behavior.