Nature 2026

Compositional Neural Control Simulator

Ref: Neural basis of compositional control (Nature 2026) ↗
2-Joint Biomechanical Effector Kinematics
End-Effector: (0.00, 0.00)
🖱 Click/Drag workspace to reposition Target Waypoint
Latent Cortical State Space (Manifold Projection)
Subspace Alignment: 0.04 (Orthogonal)
🔄 Drag to Orbit 3D Latent Space | Scroll to Zoom
Interference Index (ξ) 0.042
Kinematic Error (MSE) 0.018 m
Subspace Angle (θ) 88.4°
Null-Space Ratio (η) 94.1%
Metabolic Energy (J) 1.42
Modular Dynamical Primitives
1. Point-to-Point Reach Gain (k_reach) 1.20
2. Limit-Cycle Oscillation Amp (A_osc) 0.00
3. Obstacle Repulsion Strength (k_repel) 0.00
4. Posture Null-Space Bias (k_null) 0.40
Latent Subspace Algebra & Noise
Subspace Orthogonality Offset (Δθ) 90°
Cortical Stochastic Noise (σ_neural) 0.04
Simulation Execution Speed 1.0x
💡 Tip: Reduce orthogonality towards 0° to induce catastrophic cross-primitive interference in behavioral kinematics.
Synthetic Population Unit Raster (16 Units) M1 / PMd
Units 1-4: Reach Subspace 5-8: Limit Cycle 9-12: Obstacle 13-16: Posture
🔬 Mathematical Formulation of Compositional Neural Control
[ − Collapse ]

1. Dynamic Superposition in Modular Subspaces

The primate motor cortex (M1/PMd) avoids catastrophic cross-talk when executing compound movements by constraining individual motor sub-routines (e.g., target tracking vs rhythmic stirring) to mutually orthogonal low-dimensional latent subspaces.

ẋ(t) = ∑ᵢ Wᵢ · fᵢ(zᵢ(t)) + W_null · g(z_null) + η(t)
where ⟨U_reach, U_osc⟩ ≈ 0 (Null-Space Orthogonality)

2. Biomechanical Null-Space Projection

For redundant multi-joint effectors, intrinsic joint velocity vectors $\mathbf{\dot{q}}$ decompose into end-effector task kinematics and null-space postural self-motion via the Moore-Penrose pseudo-inverse Jacobian $J^\dagger$:

q̇ = J†(q) · ẋ_des + (I - J†(q)J(q)) · ∇_q H(q)

3. Subspace Interference & Out-Of-Distribution Robustness

When tasks are blended, the total cross-primitive interference index $\xi$ quantifies the projection of primitive dynamical flows into competing task spaces. High orthogonality ($\theta \to 90^\circ$) preserves modular compositionality.

ξ = || P_task1 · f_task2(z) ||_F / (||f_task1|| + ||f_task2||)