Inference From Inverse Dynamics in a Nonlinear Affine System
Introduction
You’re a detective. A golfer swings the club and hits a ball. You have perfect video of what happened—every position, every acceleration, every twist. The question: what forces did the golfer apply? This is inverse dynamics: detective work, working backward from what happened to figure out what must have caused it.
But here’s the detective’s dilemma: imagine you see a car accelerating downhill at constant speed. Did the driver press the gas (active force) or is the hill doing it (passive drift)? Without knowing the slope of the hill, you can’t tell. You see the acceleration, but you can’t separate the driver’s input from gravity’s free push.
The same problem haunts inverse dynamics in biomechanics. You see the club accelerating. But that acceleration comes from two modeled sources: drift (state-dependent mechanics) and applied input. Classical inverse dynamics gives you a model-consistent net load, not muscle activation, metabolic cost, or conscious intent. A drift/control split can clarify mechanical attribution while leaving those physiological quantities unidentified.
This section formalizes the limitations of classical inverse dynamics and explains how the control-affine framework together with the Zero Torque Counterfactual (ZTCF) family and Zero Velocity Counterfactual (ZVCF) solve the detective’s dilemma by separating passive from active contributions.
Classical Inverse Dynamics and Its Limitations
The classical rigid-body dynamics of the system can be written as
\[ \underbrace{M(q)\ddot{q} + C(q,\dot{q})\dot{q} + g(q) + \tau_{\mathrm{passive}}}_{\tau_{\mathrm{ID}}} = \tau_{\mathrm{input}}, \]
where \(M(q)\) is the inertia matrix, \(C(q,\dot{q})\dot{q}\) contains Coriolis and centrifugal forces, \(g(q)\) contains gravitational forces, and \(\tau_{\mathrm{passive}}\) includes passive elastic and damping forces. The left-hand side is exactly the total inverse-dynamics torque \(\tau_{\mathrm{ID}}\) — the generalized torque required to reproduce the measured motion. For the actual motion the applied (input) torque equals this total, \(\tau_{\mathrm{input}} = \tau_{\mathrm{ID}}\); that identity is what classical inverse dynamics computes.
The difficulty is not in computing \(\tau_{\mathrm{ID}}\) but in attributing it. We re-group the same total — not adding a second equation — into an inertial part and a drift part. Define the drift torque as the configuration- and velocity-dependent passive generalized force evaluated along the actual trajectory, and the inertial torque as the part carried by the mass matrix:
\[ \tau_{\mathrm{drift}}(x) := C(q,\dot{q})\dot{q} + g(q) + \tau_{\mathrm{passive}}, \qquad \tau_{\mathrm{inertial}} := M(q)\ddot{q}, \]
so that \(\tau_{\mathrm{ID}} = \tau_{\mathrm{inertial}} + \tau_{\mathrm{drift}}\) (matching the decomposition in inverse-dynamics and drift-components-wrench-double-pendulum). The active contribution is the residual after removing the drift the zero-input (ZTCF) trajectory would require:
\[ \tau_{\mathrm{input}}^{\mathrm{active}} = \tau_{\mathrm{ID}} - \tau_{\mathrm{drift}}^{\mathrm{ZTCF}}. \]
Because the drift and inertial parts are not separately observable from kinematics alone, classical inverse dynamics — which yields only the total \(\tau_{\mathrm{ID}}\) — inherently conflates passive and active contributions.
Why Classical Inverse Dynamics Cannot Identify Input Torques
The input torques satisfy
\[ \tau_{\mathrm{input}} = Bu, \]
where \(B\) maps input controls into generalized joint torques. Solving for \(u\) given only \(\tau_{\mathrm{ID}}\) would require knowing \(\tau_{\mathrm{drift}}\), because
\[ Bu = \tau_{\mathrm{ID}} - \tau_{\mathrm{drift}}. \]
However, \(\tau_{\mathrm{drift}}\) is not observable directly from kinematic measurements. Drift dynamics include inertial effects, Coriolis forces, gravitational loading, shaft deformation forces, joint passive forces, and geometric coupling terms. Many of these require full dynamic simulation or model-based estimation and cannot be separated from \(\tau_{\mathrm{input}}\) using classical inverse dynamics alone.
Thus, classical inverse dynamics cannot provide an unambiguous estimate of muscular effort in a drift-dominated system such as the golf swing. Even a model-perfect net input torque remains compatible with different muscle recruitment, co-contraction, and passive-impedance states.
Drift–Input Decomposition and Affine Structure
In the affine form, the decomposition of total generalized forces is
\[ F_{\mathrm{total}} = F_{\mathrm{drift}} + F_{\mathrm{input}}. \]
Therefore, the active input forces can be computed only if the drift forces are known:
\[ F_{\mathrm{input}} = F_{\mathrm{total}} - F_{\mathrm{drift}}. \]
Classical inverse dynamics attempts to recover \(F_{\mathrm{total}}\) but has no mechanism for computing \(F_{\mathrm{drift}}\). This is where counterfactual simulation becomes essential.
Pointwise Drift and ZTCF Trajectories
The Zero Torque Counterfactual (ZTCF) simulates the dynamics under \(u(t) \equiv 0\):
\[ \dot{x} = f(x). \]
At the branch state, the zero-input solve provides a pointwise drift-force evaluation:
\[ F_{\mathrm{drift}}(t_0) = \text{forces from the zero-input solve at }x(t_0), \]
including all inertial, gravitational, Coriolis, centrifugal, and elastic shaft contributions declared by the model. Integrating from that state produces a branched ZTCF trajectory. Once the trajectories diverge, their force difference is a counterfactual outcome difference rather than a same-state algebraic attribution.
ZVCF for Configuration-Dependent Drift
The Zero Velocity Counterfactual (ZVCF) sets all velocities to zero while keeping the configuration fixed, then evaluates the drift at this zero-velocity state. This yields configuration-dependent forces (gravity and elastic restoring forces) without velocity-dependent terms (Coriolis, centrifugal). The forces obtained from the ZVCF correspond to
\[ F_{\mathrm{ZVCF}} = f(q, 0) = F_{\mathrm{gravity}} + F_{\mathrm{elastic}}. \]
The ZVCF therefore isolates the configuration-dependent passive component of generalized forces. The pure input is obtained by subtracting the full drift, \(F_{\mathrm{input}} = F_{\mathrm{total}} - F_{\mathrm{drift}}\) (the drift the ZTCF provides at the branch instant), not by subtracting the ZVCF: \(F_{\mathrm{total}} - F_{\mathrm{ZVCF}}\) retains the velocity-dependent drift (Coriolis/centrifugal), so it equals \(F_{\mathrm{input}} + F_{\mathrm{vel.drift}}\), not input alone. In the downswing the velocity-dependent drift is large, so this distinction matters.
Hand–Club and Ground Reaction Forces as Mixed Forces
Measured hand–club forces and ground reaction forces satisfy
\[ F_{\mathrm{measured}} = F_{\mathrm{drift}} + F_{\mathrm{input}}. \]
Without drift–input decomposition, classical inverse dynamics cannot distinguish passive contributions (such as inertial loading or shaft bending) from active contributions transmitted through joint torques. Counterfactual simulation enables this separation by computing drift and input terms independently.
The proximal-to-distal hand-path study applies this distinction to the equivalent grip force reported in golf inverse dynamics. Its new numerical attribution remains fail-closed until an exact source commit and artifact hashes are pinned.
Causal Interpretation of Inverse Dynamics in Affine Systems
Because drift is the causal residue of all prior states, the generalized forces required to reproduce a motion necessarily encode the system’s history. Classical inverse dynamics at time \(t\) therefore conflates contributions from past passive dynamics with present-time active inputs. The affine decomposition clarifies this causality by separating drift from input forces, allowing torque inference to be interpreted in terms of the system’s geometry and dynamic history.
Implications for Biomechanics Research
The affine decomposition, ZTCF, and ZVCF enable: - accurate estimation of input torques,
separation of passive and active contributions to interaction forces,
interpretation of large forces without assuming high muscular effort,
reproducible inference independent of equipment or technique,
improved machine learning models based on drift-removed data. This section provides the conceptual and computational foundation for replacing classical inverse dynamics with a more rigorous framework grounded in nonlinear affine dynamics.
Limitations
The ZTCF and ZVCF framework, while powerful, operates within several constraints that practitioners should recognize:
Full model and initial conditions required: ZTCF computation requires knowledge of the complete system model—inertia matrices, damping coefficients, gravity, and shaft properties. Without precise initial conditions, the counterfactual trajectory diverges rapidly, compounding error in the estimated drift forces.
Computational cost scales with complexity: As model fidelity increases (more degrees of freedom, flexible segments, joint stiffness distributions), the cost of running high-fidelity counterfactual simulations scales superlinearly. For real-time biomechanics or online coaching, this becomes prohibitive.
Assumes perfectly known system parameters: The drift estimation is only as good as the parameterization. In practice, mass distributions, joint friction, and passive elastic properties are approximated. Small parameter errors can create large errors in drift estimation, particularly when inertial forces dwarf muscular contributions.
Cannot distinguish voluntary braking from reflexive co-contraction: If a golfer is simultaneously contracting agonist and antagonist muscles (co-contraction), the ZVCF isolation sees only the net input torque. It cannot separate deliberate stabilization efforts from reflex responses, potentially mischaracterizing the cognitive control component of motion.
<div class="laymans-terms-inner">
<p class="laymans-terms-intro">
Here is a simplified explanation of why standard methods struggle to measure golfer effort and how our new approach solves it.
</p>
<div class="laymans-item">
<h3>The Hidden Hill Problem</h3>
<p>Imagine driving a car. Your speedometer shows your speed, but it doesn't tell you how hard the engine is working. You could be speeding up because you are pressing the gas (effort) or because you are rolling down a steep hill (gravity/drift).</p>
<div class="analogy"></div>
<div class="laymans-item">
<h3>Separating the Electric Assist</h3>
<p>The golf swing is like riding an electric bike. You have your own pedaling (Input) and the motor/momentum helping you (Drift). To know how much exercise you are getting, you need to know exactly how much the bike is helping you.</p>
<div class="analogy">