Launch Monitor Vendor Reference
What This Page Is
A launch monitor turns observations of a moving club and ball into quantities that golfers can use. Understanding those quantities requires more than reading their names. Two displays labelled club path can refer to different points on the head; two closure rates can describe different rotations. A repeatable number can have an offset, and a disagreement can combine a real difference in definition with measurement error.
This reference connects vendor documentation to the mechanics and statistical questions needed to interpret it. Its purpose is to make a golf argument testable: which delivered state changed, how was that change observed, and does it explain the ball outcome? Vendor statements below describe the cited document and configuration, checked on 8 September 2026. They are not independent certification of a product’s accuracy.
Before comparing two club numbers, match the point, coordinate frame, angle convention, time or averaging window, and processing mode. Then assess agreement and uncertainty. A different convention is not automatically an error; an undocumented convention is a limit on interpretation.
The companion Launch Monitor Technology discusses sensing. Impact Mechanics and Ball Flight develops the collision and aerodynamic model. This page supplies the definitions and evidence needed to connect those models to recorded data.
Part I: Define the Measurement
I.1 Where on the Clubhead Is the Reported Point?
During a short interval, approximate the head as a rigid body. Let \(P\) and \(Q\) be two material points, with displacement \(\mathbf r=\mathbf x_Q-\mathbf x_P\) expressed in one fixed, right-handed frame. Their velocities obey
\[ \mathbf v_Q=\mathbf v_P+\boldsymbol\omega\times\mathbf r. \]
Here \(\boldsymbol\omega\) is the head’s angular velocity, not a particular projected closure rate. The formula is exact for rigid motion. Shaft flexibility can change how the head moves relative to the handle without invalidating a locally rigid-head approximation. Face deformation during collision requires a further local deformation term.
For the examples on this page, let \(x\) point right of target, \(y\) along the target, and \(z\) upward. This is a right-handed frame. Define horizontal path \(\psi\) and true velocity elevation \(\alpha\) by
\[ \psi=\operatorname{atan2}(v_x,v_y),\qquad \alpha=\operatorname{atan2}\!\left(v_z,\sqrt{v_x^2+v_y^2}\right). \]
Positive path is right of target, independently of the golfer’s handedness. The terms in-to-out and out-to-in additionally depend on handedness. Path is undefined for zero horizontal velocity; all velocity-direction angles are undefined at zero speed. A vendor’s attack angle may instead use the projection \(\operatorname{atan2}(v_z,v_y)\). That projected angle and true elevation differ when \(v_x\ne0\).
Worked point-change example. Take \(\mathbf v_P=(0,45,0)\) m/s, \(\mathbf r=(0,0.04,0)\) m and \(\boldsymbol\omega=(0,0,-50)\) rad/s. Then \(\mathbf v_Q=(2,45,0)\) m/s. The two paths are \(0^\circ\) and approximately \(2.54^\circ\). Both describe the same rigid motion. Reversing the angular velocity reverses the discrepancy; choosing \(\mathbf r\) parallel to \(\boldsymbol\omega\) makes it zero. A single offset cannot convert every golfer, club and strike.
For a small velocity change \(\delta\mathbf v\), the first-order path change is
\[ \delta\psi\approx \frac{v_y\,\delta v_x-v_x\,\delta v_y}{v_x^2+v_y^2}, \qquad \delta\mathbf v=\boldsymbol\omega\times\delta\mathbf r. \]
This exposes what a reference-point correction needs: the displacement, the angular velocity and the translational velocity. Distance between points alone is insufficient. The approximation also becomes poorly conditioned as horizontal speed approaches zero.
I.2 Measurement Includes Inference
The useful distinction is the evidence supporting an estimate, not whether an algorithm was involved. The international metrology vocabulary explicitly includes a measurement model and a measurement function: input quantities are used to obtain the reported quantity and its uncertainty. An algorithm can implement that function. See VIM 2.48 and VIM 2.49.
Camera systems estimate geometry from calibrated images, identify features, fit poses and differentiate positions. Radar systems process reflected signals and estimate motion from their temporal, spectral and spatial structure. A model can also predict an unobserved quantity using other measurements or learned relationships. These processes have different failure modes, but none becomes accurate simply by being called direct.
A schematic observation and reporting chain is
\[ \mathbf z=h(\mathbf x,\mathbf c)+\boldsymbol\epsilon, \qquad \widehat{\mathbf y}=g(\mathbf z,\widehat{\mathbf c},\mathcal M). \]
The physical state is \(\mathbf x\), calibration parameters are \(\mathbf c\), observations are \(\mathbf z\), and \(\mathcal M\) records the reconstruction model and assumptions. The measurand \(\mathbf y\) might be face orientation at a particular point and time. The same observations may support some components of \(\mathbf y\) strongly and others weakly.
For example, a line through two tracked shaft points supplies a direction, but rotation about that line leaves both points unchanged. Those two points alone cannot identify axial attitude. An off-axis feature, another sensing channel or a prior can add information. A learned prior can produce an estimate even in an ambiguous view; that is a reason to test its uncertainty and generalization, not to declare that the estimate cannot exist.
Likewise, one radial Doppler component does not identify arbitrary three-dimensional velocity or face attitude. Multiple antennas, views, time samples, shape information and motion models change what can be identified. A subscription unlocking a parameter establishes commercial availability, not which of these methods determines it.
Part II: Vendor Catalogue
The catalogue separates published definitions from evidence of performance. A feature list is not an uncertainty budget. Historical documentation remains useful, but it should not silently describe a later generation of hardware or software.
TrackMan
The 2017 club-data explanation distinguishes the geometric head center from the face center and contact location. It gives an illustrative driver path difference of roughly three degrees and discusses curved-face normals. These are context-dependent examples, not universal corrections.
That document also makes an essential timing distinction: speed, path and attack-angle reporting use preimpact motion, while face angle and dynamic loft represent orientation averaged over the collision and associated with maximum compression. Do not extend the preimpact restriction to every club parameter. The precise sampling and fitting procedure for a current product requires its own documentation.
Useful current entry points include the club-path help page, face-angle help page, and spin-loft explanation. Match the reported point and timing before interpreting a difference against an optical face measurement.
FlightScope
The data-parameter page currently defines club speed and path using the geometric center, and attack angle using the approach just before impact. It also identifies model/package availability. Its speed and acceleration profiles describe motion relative to the radar; those profiles should not automatically be read as total speed, tangential acceleration or face closure.
The club-path and face-to-path pages are additional terminology sources. A software tier can change the available outputs without changing the physical sensor. Illumination requirements for a camera-derived output can coexist with a radar channel that operates in darkness. Requirements must be attached to the specific parameter and mode.
The practical question for research is how the selected configuration maps its signals to the declared point, orientation and event, and how well that mapping performs in the intended hitting environment. Descriptions such as Fusion Tracking identify a product approach; they do not answer every calibration or uncertainty question.
Foresight Sports and Bushnell Launch Pro
Foresight’s measurement overview describes camera-based reconstruction and defines closure as heel-to-toe rotation about the shaft. Its club-data glossary distinguishes face orientation, impact location and motion. A face-plane orientation does not by itself specify the local normal everywhere on a curved driver face.
The marker guide distinguishes one- and four-marker modes and supported hardware. GCQuad, QuadMAX, Falcon and GCHawk support the expanded marker arrangement; GC3 and Launch Pro belong to the one-marker group. Therefore the full four-marker orientation/closure list must not be assigned to every device carrying related technology. Check the exact model, entitlement and manual.
Marker placement is part of the measurement procedure. A fitted marker frame, a nominal face plane and the surface normal at the actual strike are different objects. Their transformations need calibration or a shape model. This does not prove that marker placement dominates any particular experiment’s errors; that requires a sensitivity study with deliberately varied placement.
GEARS Golf
The club and ball metrics reference explicitly offers speed at the face center and at the impact point. It also offers face-angle and loft variants associated with those locations. The distinction is not limited to speed.
GEARS defines several angles through specified projections: its documented ground plane is XZ and target direction Z, so its face-on ZY attack angle is a projected angle. Its documented spin-loft construction is also projected. These conventions must be transformed before comparison with a true three-dimensional vector angle.
The same reference describes closure relative to club path. That differs from a rotation component about the shaft. Its detailed definitions are useful for model construction, but the availability of extensive outputs does not establish that GEARS has the lowest uncertainty of every commercial system. Such a ranking would require a common reference, protocol and error metric.
Uneekor
Uneekor’s current AI Studio announcement describes sticker-free Club AI configurations for EYE XR, EYE XO and EYE XO2. It also describes unmarked-ball spin tracking. Those configurations cannot be classified solely from an older QED or EYE XO sticker guide.
Start from the support resources and the exact configuration’s manual. Ball markings, club markers, camera arrangement and software entitlements are separate requirements. A claim that a current configuration needs neither kind of marker should be tested under that configuration’s stated conditions; it does not retrospectively change an older unit’s instructions.
The cited announcement supplies capabilities, not a complete mathematical definition and uncertainty budget for every output. Patent-assignee names are not a substitute for that information, and this catalogue makes no inference about ownership, licensing or shipped algorithms from corporate names.
Full Swing
Full Swing’s KIT data guide defines path using the head center before impact and lists the other displayed quantities. Its face-to-path entry is an angular difference. Some adjacent coaching descriptions simplify ball-flight relationships; for example, path alone does not determine draw or fade.
The launch-monitor patent US11844990B2 describes nonuniform antenna arrangements and learned processing of radar data. It is useful evidence that such an architecture was disclosed. It does not establish which algorithm a particular KIT firmware uses for clubface orientation, or provide an independent test of that output. Nor does the absence of a camera from a selected patent claim prove that a retail device has none.
Rapsodo: Patent and Product Scope
US11583746B2 describes a club-scanning and image-reconstruction method with radar input. Reconstruction is a normal geometric measurement operation. The patent is not a specification of the MLM2PRO, and its possible scanned-club workflow should not be assigned to that product without product documentation.
Sportsbox AI
Sportsbox connects body motion to golf analysis through video reconstruction. Its published accuracy report compares selected chest and pelvis measures with AMM3D in 30 golfers, one swing each. It reports average absolute angular differences of about two degrees at selected events. Those averages do not bound every individual’s error or validate every output, and the selected measures do not establish clubface accuracy.
Its shaft-angle explanation explicitly distinguishes shaft lean from clubhead travel and notes unmeasured shaft deformation. Even a perfectly rigid shaft’s inclination does not uniquely determine the head’s velocity: handle translation and angular motion are also needed.
The two-point ambiguity described in Part I is a geometric statement about those observations alone. It is not a complete description of Sportsbox’s proprietary model. Errors from another markerless system cannot be used as quantitative Sportsbox error bounds. Similarly, an unsuccessful database search does not establish that no validation exists.
Garmin: Approach R10 and R50
Garmin’s R50/R10 comparison explicitly identifies R50’s three-camera system and R10’s radar system. It associates R50 club reporting with the sticker point and R10 reporting with the strongest head radar return. This is a concrete example of within-vendor reference-point differences; it should not be generalized to every camera or radar product.
The R10 accuracy support page separates some radar measurements from other calculated outputs and supplies manufacturer tolerances. These remain manufacturer statements with a defined product scope, not independent coverage probabilities for every use case.
The ordinary-ball outdoor spin conditions sometimes quoted from that page must not be turned into a universal ban on short-flight spin measurement. Garmin separately documents R10 compatibility with Titleist RCT balls for indoor spin capture. Ball construction, software, environment and estimated-spin indicators belong in the session record. A possible interference warning is not proof that good spin and good club data are mutually exclusive on every shot.
Rapsodo: MLM2PRO
The MLM2PRO FAQ labels eight entries in its 15-metric list as measured, including club path and attack angle, and describes RPT balls for spin tracking. Face angle is not in that published list. Treat this as the documented list, rather than a permanent statement about every future update.
Rapsodo’s June 2025 comparison with GCQuad reports 1,021 shots, with mean absolute differences of \(1.05^\circ\) for attack angle and \(1.19^\circ\) for path. This is useful, vendor-published comparative evidence. The comparator’s uncertainty, shared alignment, sample composition and excluded shots still matter; a mean absolute difference is not a maximum error or a 95% interval.
The setup guide emphasizes alignment, leveling and the club-data hitting zone. Its feet/centimeters distance labels are not exact conversions, so consult the current device setup guidance rather than treating both printed numbers as mathematically identical.
SkyTrak, SkyTrak+ and ST MAX
SkyTrak’s current technology page describes Doppler radar for club data, a photometric camera system for ball data and machine-learning processing. It includes comparisons with TrackMan attributed to GolfBays. Thus the blanket statement that the current range publishes no hardware information is incorrect.
Keep the original optical SkyTrak, SkyTrak+ and ST MAX separate. A tolerance table for an older ball-measurement product cannot establish the newer product’s clubface uncertainty. Conversely, the presence of learned processing is not evidence of inaccuracy. Claims such as being within a specified angle of another device need the metric, shot distribution, reference setup and proportion of shots satisfying that bound before they become a usable error model.
Ernest Sports
Ernest Sports has a live support index, a Tour Plus 2.0 overview, and a retrievable historical ES Tour Plus manual. The earlier assertion that its documentation was permanently lost was wrong.
The current overview describes infrared cameras, Doppler sensors and club outputs. The historical manual lists data and operating modes, including reduced outputs in outdoor mode. This establishes documented capabilities and conditions, not complete metrological qualification. A bare percentage labelled accuracy cannot replace a parameter-specific error definition, reference and test distribution.
Swing Catalyst
Swing Catalyst is relevant because ground measurements and imported launch data are used together to interpret the swing. Its sensor-plate documentation distinguishes pressure and force hardware; the 3D Motion Plate setup guide discusses calibration and alignment. Record which hardware generated each channel. The software displaying a number need not be the instrument that measured it.
For a horizontal pressure array, let cell \(i\) carry normal force \(f_i\) at \((x_i,y_i,0)\). If the calibration converts pressure and area to force,
\[ F_z=\sum_i f_i,\qquad x_{\mathrm{COP}}=\frac{\sum_i x_i f_i}{F_z},\qquad y_{\mathrm{COP}}=\frac{\sum_i y_i f_i}{F_z}. \]
For this normal loading, \(M_x=F_z y_{\mathrm{COP}}\) and \(M_y=-F_z x_{\mathrm{COP}}\). Therefore pressure data contains more than a single total-force scalar: it locates the resultant and its horizontal moments. It does not identify horizontal shear forces or the free moment about the vertical axis. COP is undefined at zero normal force and unstable near it. A complete force plate measures additional channels; its reference origin and surface-height correction still matter. See Force Measurement Technology.
Neither pressure nor a complete net ground wrench identifies all joint torques or muscle forces by itself. Body motion, inertial parameters, other contacts and a declared inverse-dynamics model connect ground data to internal mechanics.
Part III: Where the Definitions Collide
III.1 Closure Rate Requires an Axis and an Angle
There is no universal scalar called the rotation of the club. Distinguish these quantities:
| Quantity | Mathematical Meaning | Interpretation |
|---|---|---|
| Angular-speed magnitude | \(\|\boldsymbol\omega\|\) | Total instantaneous rotation rate |
| Component about a specified shaft axis | \(\boldsymbol\omega\cdot\mathbf a\) | One component, for declared unit axis \(\mathbf a\) and relative frame |
| Projected face closure | \(\dot\phi\) | Rate of a face-normal azimuth |
| Path-relative closure | \(\dot\phi-\dot\psi\) | Rate of the unwrapped face-to-path angle |
| Closure per path length | \(d\phi/ds\) | Orientation change per distance traveled by a specified point |
The Foresight and GEARS documentation above uses different descriptions. A manufacturer’s prose still needs its precise implementation to determine whether an output is a component, an Euler-angle derivative or another relative-angle derivative. Their values cannot be converted by an assumed population-wide factor.
Let \(\mathbf n\) be a unit face normal in our fixed frame. When its horizontal projection is nonzero,
\[ \phi=\operatorname{atan2}(n_x,n_y),\qquad \dot\phi=\frac{n_y\dot n_x-n_x\dot n_y}{n_x^2+n_y^2}, \qquad \dot{\mathbf n}=\boldsymbol\omega\times\mathbf n. \]
Thus projected closure depends on both angular velocity and current face orientation. Rotation about \(\mathbf n\) does not change that normal at all. Near a vertical normal the horizontal azimuth is poorly conditioned. A shaft-axis component is a different projection, and handle twist can differ from head twist when the shaft deforms.
An expression such as
\[ \mathrm{CCV}=\mathrm{HTV}\sin\ell+\mathrm{SPV}\cos\ell \]
requires a declared decomposition of angular velocity into specified axes and a specified projected output. It is not a general identity for arbitrary proprietary closure metrics. Before using it, define every angle, sign, axis and reference frame, and include any additional angular-velocity component.
What Distance Normalization Does
For arc length \(s\) of point \(P\), with nonzero speed,
\[ \frac{d\phi}{ds}=\frac{\dot\phi}{\|\mathbf v_P\|}. \]
If the same position and orientation path is traversed with a different strictly increasing time parameter, numerator and denominator acquire the same time-scaling factor. Their ratio is unchanged. A faster actual swing can change the trajectory, shaft deformation, face orientation and timing, so this invariance does not automatically apply to it.
Degrees per second answers a temporal-sensitivity question; degrees per meter answers a spatial-sensitivity question. Both are meaningful. At fixed spatial path, shortening the time scale increases the angle error produced by a fixed clock error even when degrees per meter remains unchanged.
Why the Screw-Axis Shortcut Fails
For nonzero \(\boldsymbol\omega\), write the velocity of a point at perpendicular distance \(\rho\) from the instantaneous screw axis as a rotational component plus an axial component \(h\boldsymbol\omega\). The screw pitch \(h\) has units of length per radian. Orthogonality gives
\[ \|\mathbf v_P\|^2=\|\boldsymbol\omega\|^2(\rho^2+h^2), \qquad \frac{\|\boldsymbol\omega\|}{\|\mathbf v_P\|} =\frac{1}{\sqrt{\rho^2+h^2}}. \]
This equals \(1/\rho\) only for zero pitch and nonzero \(\rho\). For \(\rho=1\) m and \(h=2\) m/rad, the ratio is approximately \(0.447\) rad/m, not \(1\) rad/m. Substituting projected face closure for \(\|\boldsymbol\omega\|\) introduces a further error. Pure translation has no finite rotational axis to which that shortcut applies. Use the full velocity-translation equation for reference-point sensitivity.
III.2 Timing Includes an Event and a Window
First contact, maximum compression and separation are distinct events. At impact does not identify which one a system uses. Nor does a timestamp tell us whether the reported value was interpolated, extrapolated or averaged from an interval around it.
For an illustrative constant closure rate of \(2^\circ\)/ms, a \(0.2\) ms timing difference produces \(0.4^\circ\) of orientation difference. This is arithmetic under the stated rate assumption, not a typical-golfer measurement. During collision, the head can accelerate and rotate in response to contact; a precontact rate cannot be extended through the entire interval without checking that approximation.
Represent a local scalar estimate by a normalized temporal weighting function \(w\):
\[ \widehat\phi(t_*)=\int w(\tau)\phi(t_*+\tau)\,d\tau, \qquad \int w(\tau)\,d\tau=1. \]
For a smooth, locally unwrapped angle, its leading difference from the event value is
\[ \widehat\phi(t_*)-\phi(t_*) \approx \dot\phi(t_*)\mu_1+\tfrac12\ddot\phi(t_*)\mu_2, \qquad \mu_k=\int \tau^k w(\tau)\,d\tau. \]
A symmetric window can remove the first term while retaining acceleration-dependent averaging error. A small timestamp displacement contributes approximately \(\dot\phi\,\delta t\). These equations are an explanatory measurement model, not a claim about a vendor’s undisclosed filter. Large rotations need an appropriate rotation-space average rather than an arithmetic mean of wrapped angles.
Impact timing itself depends on the trajectory and the event detector. In a model with contact guard \(g(\mathbf x,t)=0\), a small state perturbation shifts the event by
\[ \delta t_*\approx -\frac{g_{\mathbf x}\,\delta\mathbf x} {g_t+g_{\mathbf x}\mathbf f}, \]
provided the denominator is nonzero and \(\delta\mathbf x\) is evaluated at the nominal event time. Near grazing contact, this linear approximation can become ill-conditioned. Synchronizing a camera and launch monitor therefore involves both clocks and a shared event definition.
III.3 Centers of Mass, Surface Points and Shape Models
A center of mass need not be visible to have a measurable velocity. If its position relative to a calibrated body frame is known, reconstructed rigid-body motion determines that velocity. What optical images alone generally do not determine is an arbitrary internal mass distribution. A manufacturer-supplied or experimentally measured center-of-mass location adds information.
Do not interchange geometric center, center of mass, face center and impact point. A head with redistributed internal mass can retain nearly the same external shape while changing its center of mass and inertia. It may then produce different collision dynamics even if a geometric-center speed is unchanged.
For a curved face, the local normal \(\mathbf n(u,v)\) depends on strike coordinates. A first-order location perturbation gives
\[ \delta\mathbf n\approx \frac{\partial\mathbf n}{\partial u}\delta u+ \frac{\partial\mathbf n}{\partial v}\delta v. \]
Bulge and roll therefore connect impact-location uncertainty to orientation uncertainty. A flat marker-plane fit is not automatically the local contact normal. Conversely, a calibrated shape model can map between them. The required correction depends on actual geometry, not a universal number of degrees per millimeter.
III.4 Build an Error Budget Before Ranking Sensors
For small perturbations, reference-point velocity uncertainty includes
\[ \delta\mathbf v_Q\approx\delta\mathbf v_P +\delta\boldsymbol\omega\times\mathbf r +\boldsymbol\omega\times\delta\mathbf r. \]
Pose error, point calibration and angular-velocity error are coupled. Differentiating noisy positions can amplify high-frequency error; smoothing can reduce that noise while changing timing and dynamic response. A high sampling rate alone specifies neither spatial resolution nor final accuracy.
For a differentiable output function with Jacobian \(J\), first-order covariance propagation is \(\Sigma_y\approx J\Sigma_zJ^T\). Keep correlations: for face-to-path \(\beta=\phi-\psi\),
\[ \operatorname{Var}(\beta)=\operatorname{Var}(\phi) +\operatorname{Var}(\psi)-2\operatorname{Cov}(\phi,\psi). \]
A shared target-line yaw offset can cancel from this difference, even while affecting both absolute angles. Different frames, filters or point definitions can prevent that cancellation. Adding independent error bars when errors are correlated can either overstate or understate uncertainty.
This is why a model-based estimate can outperform another reconstruction under a particular protocol without establishing a general superiority of radar, cameras or machine learning. Explain the observed performance and its conditions before assigning a cause.
Part IV: The Independent Validation Literature
IV.1 A Traceable Reference Comparison
Leach et al. (2017) compared TrackMan Pro IIIe and GC2+HMT against calibrated high-speed optical reconstruction: eight golfers, 240 attempted shots, three clubs, January 2015 acquisition. Ball agreement was generally stronger than club agreement. Calibration traceability and ball-specific uncertainty checks strengthen the reference; they do not remove every club-geometry or timing uncertainty.
Table 6 uses bands around the median difference, not around zero error. Selected percentages within the authors’ research bands are:
| Parameter | Half-Width | TrackMan | Foresight |
|---|---|---|---|
| Head Speed | 1 mph | 54% | 29% |
| Club Direction | \(1^\circ\) | 45% | 58% |
| Face Angle | \(1^\circ\) | 66% | 26% |
These are dispersion criteria chosen by the authors, not an industry certification or direct pass rate against a manufacturer’s absolute tolerance. The study’s median face differences were \(0.0^\circ\) and \(-0.8^\circ\), respectively.
TrackMan returned most club parameters on 62% of shots, only 19% for the wedge; wedge club data remained in pooled analysis but not individual-club analysis. Foresight returned club data on 75%. The authors noted possible interference in the simultaneous setup. These are historical protocol-specific results, not current product rankings.
A Small Statistical Counterexample
Suppose every error lies between \(4.9^\circ\) and \(5.1^\circ\). All observations lie within \(1^\circ\) of their median, yet none lies within \(1^\circ\) of zero. Reporting only the median-centered percentage would conceal the offset. Conversely, a median near zero does not rule out wide scatter. Report location, spread and the definition of the acceptance band together.
IV.2 Repeatability Includes the Golfer
Shaw et al. (2023) studied 21 low-handicap golfers across two sessions. Speed and distance measures were more repeatable than several angular measures under its protocol. Shots more than 20 yards off the intended line were repeated. The result concerns the complete player–protocol–instrument system, not isolated electronic noise.
The abstract of Bishop et al. (2023) reports high speed ICCs and lower spin-rate ICCs during indoor testing. Its journal volume is dated 2023 and its online publication 2024. The full tables were not retrieved for this review, so their unit-specific SEM values are not reproduced. A between-session change is not automatically instrument drift.
To see the distinction, write a repeated observation as
\[ y_{ij}=\mu_i+s_{ij}+b_j+\epsilon_{ij}. \]
Here \(\mu_i\) represents a golfer’s baseline, \(s_{ij}\) real shot/session variation, \(b_j\) a session-dependent instrument or setup effect, and \(\epsilon_{ij}\) other measurement error. Repeated human swings alone generally cannot separate these components. A controlled-motion reference, simultaneous independent measurements or a suitably designed variance-components experiment is needed.
In a simplified random-intercept model,
\[ \mathrm{ICC}=\frac{\sigma_{\mathrm{between}}^2} {\sigma_{\mathrm{between}}^2+\sigma_{\mathrm{within}}^2}. \]
Holding measurement noise fixed while narrowing the golfer or shot range can lower ICC. This is not a universal formula for every published ICC: one must identify the effects model, consistency versus absolute agreement, and single versus averaged measurements.
The coefficient of variation also needs care. Dividing a standard deviation by a mean angle near zero can yield a huge or undefined percentage. Changing the angular zero convention can change that CV without changing the underlying scatter. For signed path and face angles, report differences and uncertainty in degrees, with circular treatment if wrapping is relevant.
IV.3 Device Agreement Is Useful Conditional Evidence
Brennan et al. (2024) compared Mevo+ with TrackMan4 in 29 youth golfers. The abstract reports strong speed correlations but larger disagreements for some spin and angular quantities. It is evidence about those devices and conditions. The full paper was not retrieved here; the abstract’s interval terminology should not be silently relabelled or converted into a claim about the entire physically possible spin range.
Bliss and Langdown (2026) used one golfer, three clubs and an indoor setup. Of 410 attempts, 207 remained after missing-data, estimated-spin and mishit exclusions. Its tables report Mevo+ dynamic loft roughly \(7.8^\circ\) and \(9.7^\circ\) above TrackMan for the retained 7-iron and wedge shots. These are sample-specific differences, not correction constants. The study distinguishes consistency and absolute-agreement ICCs; weak agreement does not identify which device is closer to the physical measurand.
The study also judged coverage using limits constructed from the same difference distribution. That does not independently establish practical equivalence. Acceptable error in degrees, speed or distance must be set from the intended decision, separately from the observed distribution. A wide interval can contain about 95% of observations and still be unusable for a one-degree coaching question.
For paired measurements of the same quantity,
\[ y_A=x+b_A+e_A,\qquad y_B=x+b_B+e_B, \]
so their difference identifies \(b_A-b_B+e_A-e_B\), with
\[ \operatorname{Var}(y_A-y_B)=\sigma_A^2+\sigma_B^2 -2\operatorname{Cov}(e_A,e_B). \]
Shared error can make agreement look excellent while both systems are wrong in the same direction. Independent error in a good comparator can contribute to the observed difference. If definitions differ, their deterministic difference enters this expression too.
Correlation alone cannot resolve this: \(y_B=y_A+5^\circ\) has perfect correlation across a varying sample and a five-degree offset. A comparative study is still useful validation evidence when the criterion and intended use are justified. Calling every device comparison not validation discards information; calling agreement absolute accuracy overstates it.
IV.4 What a Qualification Study Should Report
The studies above do not qualify every current model, firmware, club, ball or environment. This review does not establish an exhaustive list of devices with no published validation. An inaccessible paper or missing search result is a retrieval boundary, not proof of nonexistence.
A useful qualification report should include:
- The measurand: point, frame, sign, angle construction, units, event and temporal window.
- The configuration: hardware, firmware, software, club geometry, ball type, markings, lighting, flight length and target alignment.
- The reference: calibration chain, reference uncertainty and how point/time conventions were matched.
- The sample: golfers or robot motions, clubs, speeds, strike locations, handedness and independent repetitions.
- The missing data: attempted shots, returned parameters, estimated flags and every exclusion rule, separated by condition.
- The errors: signed bias, distribution, absolute differences, limits with uncertainty and dependence on the quantity’s magnitude.
- The decision criterion: acceptable error chosen before seeing the results, tied to the intended experiment or coaching decision.
For approximately stationary, normally distributed paired differences, mean \(\pm1.96\) standard deviations describes approximate limits of agreement. Those limits are not confidence intervals for the mean; their own estimates have uncertainty. Repeated shots from one golfer are not independent samples of different golfers, and proportional bias or heteroscedasticity can require a different model.
Selection is part of performance. Let \(R=1\) mean that a usable result is returned. The error distribution conditional on \(R=1\) need not represent all attempted shots. If difficult strikes are disproportionately rejected, excellent retained-shot agreement can coexist with poor coverage of exactly the shots a golfer wants to diagnose. Conversely, a declared abstention can be preferable to a confidently wrong value. Report both availability and conditional error.
Part V: From Delivery Numbers to a Swing Argument
A slice is a ball-flight outcome, not a unique club-path value or a unique body-motion fault. Launch direction, spin vector, speed, strike location and environmental forces all contribute to the observed flight. For a fixed club speed, changing delivered orientation or strike location can change the result without the same change in head-center path.
For a velocity elevation \(\alpha\), face-normal elevation \(\lambda\), path \(\psi\) and face azimuth \(\phi\) in our common frame, the angle \(\gamma\) between velocity and normal satisfies
\[ \cos\gamma= \cos\alpha\cos\lambda\cos(\phi-\psi) +\sin\alpha\sin\lambda. \]
For matching horizontal directions, this reduces to the appropriate magnitude of \(\lambda-\alpha\). With different horizontal directions it generally does not. Projected loft-minus-attack is therefore not a general three-dimensional spin-loft identity. This geometry also does not, by itself, predict the complete ball spin: friction, strike offset, inertia, deformation and contact evolution enter the collision model.
A reported face-to-path angle remains a well-defined difference under declared conventions even if path refers to the head center and face orientation to the contact location. To use it in contact mechanics, map the velocity to the relevant contact point and choose a consistent surface normal and event. This is a modeling step, not grounds for declaring the display meaningless.
Consider the earlier point-change example. Coaching a displayed path from \(2^\circ\) to \(0^\circ\) does not specify what happened to face angle, contact location, angular velocity or speed. It therefore cannot prove that zeroing the display causes a slice. A causal claim would require an intervention that measures those coupled changes and compares actual outcomes.
Population Benchmarks Are Descriptions
TrackMan does publish amateur averages: its launch-angle page and swing-plane page include male and female amateur Combine groups. The earlier assertion that only tour averages exist was false. Those pages do not establish a representative population distribution of amateur club path or face angle.
An average is not an optimal target for every golfer. Specify the sampled population, clubs, task, shot selection and averaging method. A player can improve dispersion while moving away from a population mean, or match a mean while retaining a harmful variability pattern. Similarly, a vendor’s example of yards of curvature per degree is conditional on ball flight and cannot be a universal conversion.
Connect the Instruments to the Mechanical Model
The measurement chain is part of the swing model. Body motion and ground contact constrain the evolving mechanical state; the grip and flexible shaft transmit forces and motion; the delivered head state and contact geometry determine the collision; the resulting ball state drives flight. Each link requires its own observations and assumptions.
Launch data strongly constrains an outcome, but it generally cannot identify a unique joint-torque history or muscle strategy. Several coordinated motions can produce similar delivery. Ground-force data alone does not determine how mechanical work is distributed internally. A body-motion reconstruction and an inertial model help close the equations, but inverse-dynamics results still depend on differentiation, external contacts, parameter estimates and model completeness.
For a model \(\dot{\mathbf x}=f(\mathbf x,\mathbf u)\) and recorded output \(\mathbf y=h(\mathbf x)\), the sensitivity \(Dh\) determines which local state changes affect the recorded quantities. Directions in its nullspace can remain invisible to that measurement at an instant. This is why matching club speed does not establish matching motion, energy transfer or control. A fuller time series and additional independent channels may distinguish states that one impact display cannot.
An effective experiment starts with a specific question, such as whether a change in handle motion alters contact-point velocity through shaft deformation. It then records synchronized handle and head motion, relevant forces, strike location and ball launch, and propagates measurement uncertainty through the model. The claim should survive plausible changes in point mapping, timing and filtering before being interpreted as a mechanical mechanism.
Part VI: Use the Evidence to Make Better Decisions
For a practice session, keep the configuration and alignment consistent, retain shot-level data and flags, and compare changes with the variability of that protocol. For cross-device comparison, translate definitions first and use simultaneous shots where possible. For research, qualify the complete measurement chain over the conditions that the argument actually uses.
The unresolved questions are specific: the current model’s reference point, its event and filter, its error distribution, its non-return behavior and its performance outside the tested sample. These are more useful requests to a vendor or investigator than an undifferentiated demand for an accuracy percentage.
The aim is neither to dismiss launch monitors nor to let their labels stand in for mechanics. It is to connect a recorded number to a physical quantity, a defensible uncertainty statement and a question about the coupled swing that the available evidence can answer.
References
Vendor links identify the specific documents used above. The bibliography records the research studies; full-text and abstract-only reading limits are stated where they affect the interpretation.
References
Citation
@online{olson2026,
author = {Olson, Dieter},
title = {Launch {Monitor} {Vendor} {Reference}},
date = {2026-08-03},
url = {https://affinedrift.com/articles/launch-monitor-vendor-reference.html},
langid = {en}
}