Lessons From Historical Fits
Evidence snapshot: all repository citations pin to commit d7c7d07add0d7407a5a3376d5fe4a97aee9112a0. The numerical results below are transcribed from checked-in handoffs and receipts. They were not rerun for this chapter. Each result is labelled by what its evidence actually establishes. The purpose is to extract reusable lessons, including failures, rather than to rank methods by their best isolated number.
Forward-dynamics matching is experimental science as much as optimization. A fit is conditional on the capture, target markers, coordinate frame, model topology, body geometry, initial pose and velocity, control basis, integrator, contact model, observation mask, and acceptance metric. Changing one can change the apparent success. The repository’s historical runs make this concrete: a very accurate partial club trajectory can coexist with poor distal body markers; an optimizer can remove bound saturation and expose a structural model floor; a saved pose can look plausible while its support contact collapses; and a process that exits successfully can still fail every physical gate.
1. Start With the Evidence Labels
The most important lesson from the historical record is to name the driving mode. UpstreamDrift contains inverse-kinematic exports, feedback-assisted holds, prescribed-coordinate experiments, fitted torque candidates, and independent torque-driven replay. These answer different questions. An IK clip answers whether a chosen configuration can resemble target frames. A feedback-assisted hold asks whether a controller can stabilize a pose under given forces and support. A torque-driven replay asks whether one set of controls, from one initial state and without target-state resets, generates the measured trajectory. Only the last can support a forward-dynamics matching claim, and even it needs contact, force, continuity and model validity checks.
The Simscape refined-Human reference states the boundary plainly: delivered clips are IK and dynamics-unqualified, feedback-assisted fixed holds do not establish full-swing dynamics, and full Human forward dynamics remains unqualified. This is a central finding, not an editorial caveat. The clips are useful for visual inspection, frame mapping and pose residuals, while the open dynamic gates identify the work still required. The maintained status is in the Simscape matching reference README; the calculation-level source is its linked simscape_matching_reference.tex.
2. Tour-Average Simscape: Short Horizons, Then Continuation
The tour-average effort demonstrates why forward validation should be incremental. The maintained tour-matching README records that 22 native tests passed in MATLAB R2025b Update 5 and that a 20 ms model simulation completed with finite states and club/grip positions. The same document records a later 0.4 s quadratic forward-dynamics candidate independently replayed with 3.119775 mm marker RMS, while warning that full-swing fitting was still unqualified at that checkpoint. A short-horizon match establishes local fit feasibility; it does not establish a full downswing, impact or follow-through.
The capture audit in that guide matters as much as the optimizer. The driver file has 654 frames at 360 Hz, source units in metres and no event group. Some marker tracks have substantially fewer valid samples, and an unlabeled cluster cannot be silently relabelled as a grip or clubhead. The guide also warns that skin markers are not anatomical joint centers and that the model’s output extractor can construct a time grid or fill unavailable signals with NaNs. Accordingly, accepted simulation data must preserve actual solver timestamps, gaps and marker validity. A convenient array shape is not data provenance.
The incremental fitting protocol hashes the capture, SLX dependencies, input MAT and marker map; first qualifies registration and fixed geometry; and then increases the fitted time prefix while retaining all earlier observations. Each prefix is reintegrated from the same initial state. This avoids a hidden reset at the stitch point and makes earlier trajectory regression measurable. If a single global sixth-order polynomial cannot fit the complete interval, the protocol proposes a piecewise representation with explicit continuity constraints and a newly qualified native implementation. The final gate is a cold native replay with tighter tolerances and no kinematic prescription. The guide distinguishes the weighted optimization objective from unweighted physical-distance metrics and asks for per-marker and per-phase error, continuity, torque and velocity limits, failures, and sensitivity.
A later two-window direct-node result reaches 0.85 seconds of uninterrupted R2025b replay. The run 102 handoff reports whole-marker RMS 20.267 mm, early RMS 9.995 mm, club-cluster RMS 8.389 mm, pelvis-yaw error 0.610 percent, zero active coefficient or node bounds, and strong Pinocchio/Simscape parity. These are notable improvements over its cited preceding runs. Yet the exact Simscape replay qualification JSON says is_qualified: false. The terminal RMS is 40.303 mm versus a 35 mm gate, despite early and club gates passing. The receipt says contact-force and penetration measurements were unavailable. This is the right interpretation: a high-quality partial dynamic result with an explicit failed gate and missing physical quantities.
The run teaches three optimization lessons. First, active bounds can be numerical artifacts: run 101 had 22 controls and 11 node coordinates at bounds; recentering control bounds and anchoring the shooting chart at the integrated endpoint removed active bounds in run 102. A solver stuck at a box edge may need better coordinates, not a wider physical model. Second, global mean error can hide terminal failure; acceptance needs phase-specific and marker-family metrics. Third, cross-engine agreement verifies the equations and implementation only to the tested model contract; both engines can agree on the same incomplete topology or absent contact measurement.
3. Human Ellipsoid Experiments: Geometry and Observability
The Human ellipsoid research history shows why geometry, orientation, mass and contact cannot be treated as cosmetic switches. The maintained Simscape reference compares tour-average Capture A and owner Capture O as separate sources. It warns that appearance does not establish skill, and that marker-protocol differences, offsets and geometry can explain apparent posture distortion. Refinement to subject-specific longitudinal dimensions improved geometric fit, but transverse artistic proportions, head geometry and shoe size remained fixed. A native subject-physics adapter applied declared mass and selected lengths, but the selected IK videos continued to use baseline physical parameters. Thus visual adaptation did not amount to a subject-specific dynamic model.
Position-only fits exposed orientation ambiguity: on three sample frames, position RMS was around 12–13 mm while the left-foot angular discrepancy approached 165–167 degrees. Adding an orientation objective at one tested low weight reduced angular error but worsened position RMS to roughly 39–51 mm, and that candidate was rejected. A fresh calibration on those same frames subsequently achieved approximately 1.2 mm position error and sub-degree foot angles, but the record correctly says those were calibration—not held-out validation—frames. The experiments show both that orientation data can resolve a weakly observed direction and that a selected weight can disturb other terms. They do not establish that a lower training-frame residual generalizes.
The body topology audit found 37 independent fit coordinates, correcting earlier hard-coded counts and indexing. That correction is an important model-contract repair: optimization over wrong dimensions can report failures or silently map residuals to the wrong joints. The same experiments identified contact assumptions as a separate problem. Treating a flat sole at address as an orientation calibration is not a measurement of a ground reaction force. A read-only FK audit found contact-sphere penetrations relative to the ground plane, but this was geometry evidence only; it was not a simulated force measurement and no dynamics were run. The reference warns that ground placement, gravity and consistent dynamic initialization require separate validation.
The supported-hold campaign supplies especially useful negative evidence. Balance-enabled tour and owner holds moved the pelvis about 20.0 and 14.8 mm and rotated about 5.6 and 4.4 degrees, exceeding 5 mm and 1-degree pose limits. Increased stiffness reduced movement but exceeded a peak-force limit. Ramped feedforward improved force gates in tested windows but still failed pose gates. Longer stance simulations lost normal force, tilted substantially and dropped the pelvis below the plane; increased damping alone also failed. These tests demonstrate that a plausible initial pose, strong servo and clean exit status do not establish stable support physics. The associated source and receipts retain rejected experiments and state that the full Human swing remains unqualified.
4. Historical Neural Fits: Code and Experiment Are Different Objects
Older design records proposed four alternatives: repeated native optimization, a learned forward map with differentiable inversion, a direct inverse CVAE, and a Python-to-Simscape bridge. Current source review shows a mixed but more mature landscape than these early READMEs suggest. Shared PyTorch FiLM forward surrogate, compact model, per-step dynamics model, CVAE, deterministic regressor, timestep inverse-dynamics model, and masked proposals all have substantive model/training code and unit tests. Legacy option2_nn_surrogate/surrogate.py and option3_inverse_nn/inverse_cvae.py remain stubs, but they are not the current shared implementations.
The experimental record nevertheless does not justify broad acceleration claims. The NM-00 inventory marks two conventional checkpoint paths absent; it quarantines unsupported mean-baseline training claims where no curves or checkpoint hash are available. NM-04’s trainer tests do not imply a native teacher corpus. NM-05’s baseline tests do not imply successful training. NM-06 says synthetic fixtures and no native training success. NM-07 is explicitly diagnostic because earlier ablation numbers were hard-coded, not measured. NM-08’s inference contracts use synthetic fixtures. NM-09’s checkpoint matrix is diagnostic and has no native ODE replay qualification on reduced planar mechanisms. NM-10’s historical speed receipt calls its costs and baseline estimates unmeasured. The corresponding receipts are linked from docs/plans/neural_motion_matching/evidence.
There is an internal inconsistency: the later model-card turnover states measured speedups and lists several models as promoted, while the artifact and validation receipts retain diagnostic, synthetic-only or unmeasured status. A textbook must cite both and decline to repeat promoted speed values until source-hashed model weights and run logs reconcile the statuses. The lesson is not to discard neural approaches; it is to make artifact lineage a hard scientific interface. A model-card verdict must be derived from immutable checkpoint, train data/split, command, runtime, candidate, replay and benchmark receipts, rather than static prose or an estimate.
The neural ablation findings still guide method design qualitatively. A forward surrogate can offer smooth gradients, but optimization can exploit inaccurate directions and propose nonphysical states; trust regions and comparison of surrogate to native directional derivatives are necessary. A hybrid that uses the learned model to initialize an exact/native polish step is safer. A physics-structured residual can preserve an analytical rigid-body prior and bound the learned correction. Diffusion or repeated sampling is useful when multimodality matters, but may not meet interactive latency. These are design conclusions recorded in a diagnostic receipt; they are hypotheses and safeguards, not quantified performance findings.
5. From Historical Failures to Adaptive Model Constraints
The most promising adaptive-dimension idea must respond to the confounds identified above. A trajectory residual alone cannot decide whether to alter segment length, marker attachment, camera offset, initial pose, inertia, actuator gain or torque basis. Free joint optimization is especially dangerous: extra model dimensions can absorb a wrong coordinate map or unmodeled marker motion and appear to “improve” a fit. Conversely, a fixed model can hit a structural error floor even after optimizer bounds are no longer active. Run 102’s head/back cluster residual provides an example: its handoff attributes a residual floor to a topology without an independent neck coordinate. The consequence is not “add any degree of freedom”; it is to formulate a topology-change hypothesis and test it against held-out head markers, orientation, dynamic replay, contact and model-complexity cost.
A robust adaptive model should learn only low-dimensional, physically interpretable shape factors. Let the base segment lengths be l0; learn l = l0 ⊙ exp(Bz) where z is a small latent and columns of B encode permitted proportional changes, such as shared upper-limb scale or leg-to-trunk ratio. Put anthropometric priors on z, enforce positive lengths and mass/inertia consistency, and preserve topology. Marker attachments should be separate parameters with tighter priors. Camera extrinsics should be calibrated independently when possible. Control coefficients should use a normalized time basis and physically interpretable bounds. This variable partition makes compensation visible.
Fit in stages: calibration and marker validity; geometry and initial state; control fit with model dimensions fixed; only then a restricted joint update if sensitivity analysis shows geometry is observable. Compute a scaled Jacobian of held-out marker/orientation residuals with respect to each parameter block. Use singular values and profile fits to find confounded directions. If a parameter changes little in the data likelihood or trades almost perfectly against torque coefficients, label it unobserved and keep its prior. Confidence should reflect empirical parameter coverage and calibration, not subjective probability that a reconstruction is a real golfer.
Validation should use more than random frame splits. Adjacent frames leak nearly identical poses. Hold out complete capture trials, view angles, movement phases or marker groups. For a single capture, hold out contiguous time windows and report extrapolation separately. Evaluate both the fit target and measurements excluded from fitting. A dimension model that lowers fit-frame RMS but worsens a held-out segment or fails contact is rejected. On small data, prioritize synthetic truth-recovery for implementation correctness, then native simulator-generated trials, then real-capture evaluation with honest uncertainty.
6. An Evidence-Preserving Research Sequence
The historical record suggests this order of work. First freeze the source capture, event interpretation, frame transformations, units, masks and marker attachments. Then establish a geometric residual floor using a qualified model and held-out targets. Confirm model topology, initial state and native actuator/control path. Next, run direct native fits with normalized bases and repeatable restarts; save every failed candidate. If an optimizer saturates, examine coordinate conditioning and sensitivity before widening bounds. Introduce surrogate/inverse networks only after a versioned native dataset and group-based split exist. Require calibrated OOD rejection and native forward polish. Finally, cold replay the selected candidate from the original initial state with solver settings stricter than the fit. Evaluate each phase, marker group, orientation, contact, torque, joint range, continuity and energy consistency, and publish failures and unavailable quantities.
This sequence converts historical “best fit” stories into reproducible evidence. A report can then distinguish four outputs: (1) kinematic visualization candidate; (2) dynamic tracking solution with feedback or transcription constraints; (3) independent forward replay that passes declared numerical gates; and (4) physically qualified model fit with contact and force evidence. Keep the status attached to the artifact, not only the prose. If a later geometry, control basis, marker map, model file or runtime changes, invalidate the old qualification and rerun the relevant gates.
The historical fits show a field that has made real progress: native execution and cross-engine parity, model-coordinate corrections, an expanding cost and optimizer toolkit, carefully preserved negative experiments, and receipts that separate accepted from rejected gates. They also show why the central research question remains open. A low trajectory error is only one measurement of a coupled system. The next method should improve identifiability and evidence quality before it tries to reduce that number further.
Citation
@online{untitled,
author = {},
url = {https://affinedrift.com/articles/forward_dynamics_matching/03_historical_lessons.html},
langid = {en}
}