Tyre and rider effects on bicycle stability

The authors got their eigenvalues by fitting decaying oscillations in a non-linear time-domain simulation, so wherever a mode grows instead of decaying the simulated bicycle falls over before there is enough signal to fit, and the paper simply leaves those speeds blank, including everything below 4 m/s.

Question

Can weave and capsize eigenvalues of this multibody bicycle model be obtained across the whole speed range, unstable regions included, to better than 0.2 m/s in weave speed?

Hypothesis

Direct numerical linearisation of the same multibody model recovers the published benchmark eigenvalues where time-domain identification fails, to within a weave-speed resolution of 0.2 m/s.

Attempt type

computational

Design

Three methods are run on one model over a speed grid from 1 to 12 m/s. Arm A is the paper's own time-domain fit to a perturbed non-linear run, as the reproduction arm. Arm B linearises the same multibody model about straight running at each speed, by finite-difference Jacobian of the state equations, and takes eigenvalues directly, which works whether a mode grows or decays. Arm C identifies the model while a proportional-derivative roll-feedback steer torque holds it upright, then divides out the known controller to recover the open-loop poles. Each arm is run twice: once on the Carvallo-Whipple benchmark with knife-edge wheels, where published eigenvalues give a known answer, and once with the Magic Formula tyre. The measure is absolute error in the real and imaginary parts of the weave and capsize eigenvalues against the benchmark table, plotted against speed, plus the weave speed each arm reports.

Controls

• known-positive-task All three arms run on the Carvallo-Whipple benchmark bicycle of Meijaard et al. 2007, whose eigenvalues are tabulated, including the low-speed region where weave is unstable.
  Rules out: That a method which returns nothing in the unstable region is failing because the dynamics are genuinely unidentifiable rather than because the method is wrong, since here the true answer is already known.
  Vocabulary note: In this field the known-positive task is a published parameter set with published eigenvalues, not a positive control reagent.
• background An unperturbed run at each speed, with zero initial lateral disturbance, passed through the same identification pipeline.
  Rules out: That eigenvalues reported near the stability boundary are artefacts of the solver's own numerical noise being fitted as a mode.
  Vocabulary note: Used here as a numerical noise floor, not a biological background signal.
• within-subject One model file, one tyre parameter set and one set of rider parameters shared by all three arms, with only the eigenvalue extraction method changing.
  Rules out: That differences between arms come from differences in the model rather than from the extraction method.

Sensitivity

The target resolution can be set from the record: the smallest weave-speed difference the paper reports as real is 0.2 m/s, for the thinner tyre 3, so any method that cannot pin the weave speed to better than 0.2 m/s cannot support the paper's own smallest claim. A tolerance on the eigenvalue real part cannot be set from the record, because the numeric precision of the published benchmark eigenvalues is not quoted anywhere in the graph JSON and the paper reports no agreement statistic for its 4 to 10 m/s match. Setting that tolerance needs the benchmark table itself and a stated acceptance band.

Would count as

Supported → positive · opposed → negative-not-achievable
Inconclusive if: The reproduction arm fails to reproduce the paper's published 4 to 10 m/s agreement, because the ADAMS model file and tyre property files are not published, so a rebuilt model would differ and there would be no anchor against which to judge the other two arms.

Already done

• The linearised equations of motion of the benchmark bicycle and its eigenvalues over forward speed are published as a reference table, with weave and capsize identified. (Meijaard, Papadopoulos, Ruina and Schwab 2007, Proceedings of the Royal Society A, doi:10.1098/rspa.2007.1857)
  So: Supplies the known-answer control at every speed including the unstable region, which is exactly where the paper's method goes blank.
• Eigenvalues of a real riderless bicycle were extracted from measured roll rate, yaw rate and steer angle by curve fitting, and the model was found fairly accurate only for low-speed low-frequency behaviour. (Kooijman, Schwab and Meijaard 2007, Multibody System Dynamics, doi:10.1007/s11044-007-9050-x)
  So: Shows the same curve-fitting limitation arises with measured data, so the fix has to be a method change rather than a longer run.
• Reduced linearised equations of motion for a single-track vehicle with toroidal wheels and suspensions were obtained as explicit functions of the design parameters, and eigenvalues taken directly from them. (Agundez, Garcia-Vallejo and Freire 2024, Multibody System Dynamics, doi:10.1007/s11044-024-09974-4)
  So: Demonstrates that arm B is already standard practice for this vehicle class, which makes the design cheap and lowers the novelty of the method itself.
• An open-source modular framework builds Carvallo-Whipple bicycle and rider models symbolically, including moving rider arms driven by elbow torques, and exposes the equations of motion. (Stienstra, Brockie and Moore 2023, The Evolving Scholar, doi:10.59490/660179a06bf1082286458109)
  So: Gives a ready implementation route for arm B that includes the passive rider arms this paper cares about, without an ADAMS licence.
• Eigenvalues of weave, capsize and wobble were computed over a city-bicycle speed range from analytical models of increasing complexity including tyre properties and frame compliance, and with tyre sliding included the capsize mode is always stable. (Doria, Roa and Munoz 2019, Mechanical Sciences, doi:10.5194/ms-10-229-2019)
  So: Independently reproduces the paper's capsize result by direct linearisation, which is a partial answer already in the literature and a cross-check for arm B.

Searched: Searched OpenAlex for the paper's own 31 resolved references, then for post-2015 work on experimental validation of the Whipple bicycle model, on linearisation of multibody bicycle models with tyre contact, and on eigenvalue estimation for unstable linear systems from short records; searched the web for bicycle weave and capsize eigenvalues extracted by curve fitting. I found no paper that specifically treats the problem of identifying eigenvalues of an unstable single-track vehicle mode from a diverging time record, which is why the design routes around it by linearising instead.

Feasibility

A multibody vehicle dynamics group with access to the ADAMS model or willing to rebuild it in a symbolic framework, for example the groups at Twente, Padova, Delft or Seville · 3 to 6 months · low, because it is entirely numerical and the only real cost is software licences and one researcher's time
Blockers: The ADAMS model file and tyre property files are not published, so arm A has to be rebuilt from the paper's description; The measured tyre coefficients come from Doria et al. and are not available as raw data; The Magic Formula tyre with relaxation adds states, so the linearised system has more eigenvalues than four and the weave and capsize labels have to be re-established rather than assumed

Strongest objection

A specialist would say the premise is a self-inflicted wound: nobody in single-track vehicle dynamics identifies eigenvalues from a non-linear time-domain run when the equations can be linearised, so the experiment confirms something the field already assumes, and its only real content is whether the ADAMS tyre and contact formulation can be linearised cleanly at all. The sharper objection is that with a relaxation-length tyre the mode set changes, so arm B and arm A are not estimating the same quantities and the error curves may not be comparable.

Not derivable from what is published

• The perturbation magnitude, type and simulation length used in the paper's identification
• The numeric precision of the published benchmark eigenvalues, and therefore an acceptance tolerance on the eigenvalue real part
• Whether the ADAMS contact and tyre formulation supports direct linearisation in the version the authors used
• The model file and tyre property files themselves

The tyre measurements behind this model covered sideslip and camber but not driving or braking, so the longitudinal force against slip curve and a longitudinal slip stiffness of 4800 N were borrowed from motorcycle tyre data, and the authors flag this as an untested assumption rather than as a plan.

Question

Does replacing the assumed 4800 N motorcycle longitudinal slip stiffness with a value fitted to measured bicycle tyre data change the predicted weave speed of the full model?

Hypothesis

A longitudinal slip stiffness fitted to measured bicycle tyre longitudinal force against slip data shifts the predicted weave speed by at least 0.2 m/s relative to the assumed 4800 N motorcycle value.

Attempt type

computational
Vocabulary note: The first block is a re-analysis of someone else's published bench measurements, so it sits between computational and evidence-synthesis; if new rig measurements are added it becomes a physical bench experiment, for which the vocabulary offers only wet-lab.

Design

Block one reuses published measurements. Fit a Magic Formula longitudinal slip curve to the measured longitudinal force against slip data of Maier et al. 2018 for a trekking bicycle tyre, extract the slip stiffness at the wheel loads of the simulated 80 kg rider, and substitute it for 4800 N in the full model. Arms are the published value, the fitted value, and a one-at-a-time bracket at 50 percent and 200 percent of nominal, following the paper's own protocol. Block two adds the only manoeuvre in which longitudinal slip is non-zero: a straight run at constant mild deceleration, with the same three parameter arms, since straight running at constant speed carries almost no longitudinal slip. The measure in both blocks is weave speed and the speed at which capsize re-stabilises.

Controls

• known-positive-task Before changing anything, reproduce the published case 6 result with the as-published parameters: weave speed 9.3 m/s against the case 1 baseline of 4.9 m/s, and a stable capsize mode.
  Rules out: That a null result comes from a broken or mis-parameterised rebuild of the model rather than from the longitudinal stiffness genuinely not mattering.
  Vocabulary note: Again a published numeric result standing in for a positive control.
• within-subject One bicycle, one rider parameter set and one tyre coefficient set throughout, with only the longitudinal parameter changed, which is the paper's own one-at-a-time protocol.
  Rules out: That the weave-speed change is caused by some other coefficient moving at the same time.
• no-treatment A constant-speed straight run with drive and brake torque set to zero, so longitudinal slip is nominally zero, run with all three parameter arms.
  Rules out: That any difference seen in the braking block is an effect of the braking manoeuvre itself rather than of the longitudinal parameter.
  Vocabulary note: Here no-treatment means a manoeuvre with no drive or brake torque, not an untreated subject.

Sensitivity

Two thresholds come straight out of the record. The smallest weave-speed difference the paper calls real is 0.2 m/s, so that is the detection target. The paper's own one-at-a-time study also gives the scale of a negligible coefficient: doubling the cornering stiffness changed weave speed by under 1 percent, and the authors treat that as no effect, while doubling the twisting torque coefficient changed it by about 25 percent. So if doubling the longitudinal slip stiffness moves weave speed by under 1 percent, this is a real null on the paper's own terms. What cannot be derived is the uncertainty on the fitted slip stiffness itself, which needs the scatter in the Maier et al. curves.

Would count as

Supported → positive · opposed → negative-no-effect
Inconclusive if: The measured longitudinal data come from a trekking tyre at loads and pressures that do not overlap the four tyres and the 2 to 5 bar range used in the paper, so the fitted stiffness is an extrapolation and a null cannot be separated from a mismatch of specimens.

Already done

• Vertical and longitudinal characteristics of a typical trekking bicycle tyre were measured on an inner drum test bench, giving characteristic curves of longitudinal force against slip with road condition, wheel load, speed and inflation pressure varied. (Maier, Hillenbrand, Wrede, Freund and Gauterin 2018, Tire Science and Technology 46(3):153-173, doi:10.2346/tire.18.460301)
  So: The measurement the authors said was missing now exists for at least one bicycle tyre, so block one needs no new rig time and the design becomes a cheap substitution test.
• A bicycle tyre test rig was built that measures vertical force, lateral force, self-aligning torque and the moment about the longitudinal axis, can measure longitudinal force after rearrangement, and can take electric motors to produce longitudinal slip. (Ballo, Gobbi, Mastinu and Previati 2020, ASME IDETC, DETC2020-22141, doi:10.1115/detc2020-22141)
  So: Names the hardware that could extend the measurement to the specific tyres in this paper if the substitution test shows the parameter matters.
• Motorcycle tyre measurements and Magic Formula models, including longitudinal behaviour, were published and became the standard parameter source for single-track vehicle simulation. (de Vries and Pacejka 1998, Vehicle System Dynamics, doi:10.1080/00423119808969565)
  So: This is the source class the borrowed 4800 N belongs to and it is in this paper's reference list, so it is the comparator arm, although the record does not confirm it as the specific origin of that number.
• Sideslip and camber characteristics of bicycle tyres were measured and the mechanical properties identified for bicycle dynamics modelling, covering lateral force, self-aligning torque and camber but not driving or braking. (Doria, Taraborrelli and others 2011 and 2012, Vehicle System Dynamics, doi:10.1080/00423114.2011.615408 and doi:10.1080/00423114.2012.754048)
  So: Confirms the gap is real in the source dataset, and fixes which coefficients must be held constant while the longitudinal one is varied.
• Internal pressure, vertical force, speed and rolling surface temperature all affect the measured mechanical characteristics of a racing bicycle tyre, with speed mattering mainly below 1.5 degrees slip angle and surface temperature dominating above 3 degrees. (Dell'Orto, Ballo, Mastinu, Gobbi and Magnani 2023, European Journal of Mechanics A/Solids, doi:10.1016/j.euromechsol.2023.105010)
  So: Warns that a longitudinal stiffness lifted from one rig at one temperature may not transfer, which is a limit on how much weight the fitted arm can carry.

Searched: Searched OpenAlex for bicycle tyre longitudinal slip force measurement, bicycle braking tyre longitudinal force, and bicycle tyre test rigs, then searched Crossref within Tire Science and Technology and across all journals for longitudinal force against slip for bicycle tyres, and searched the web for published Fx against kappa bicycle datasets. Apart from Maier et al. 2018 I found no openly published bicycle tyre longitudinal force against slip dataset: the Politecnico di Milano rig can be arranged to measure it but the published campaigns from that rig cover lateral force and self-aligning torque only.

Feasibility

A multibody group for block one working from the Maier et al. published curves; a tyre bench group such as Politecnico di Milano or Karlsruhe for any new longitudinal measurement · 1 to 3 months for block one, 6 to 12 months if new rig measurements are needed · low for block one because it reuses published measurements, moderate if the rig campaign is run
Blockers: The Maier et al. curves are published as figures in a paywalled journal, so digitising them introduces an error the record cannot bound; The trekking tyre measured is not one of the four tyres in this paper; The paper does not say which motorcycle source the 4800 N came from, so the comparator arm is reconstructed rather than copied

Strongest objection

A specialist would say block one is guaranteed to find nothing and everybody knows why: in straight running at constant speed the longitudinal slip is essentially zero, so the longitudinal force derivative does not enter the linearised lateral dynamics, and the authors listed this as a limitation for completeness rather than as a live question. The objection to block two is that once there is braking the answer is governed by the Magic Formula combined-slip weighting functions and by load transfer onto the front wheel, not by the longitudinal slip stiffness alone, so attributing any effect to that single parameter is not defensible.

Not derivable from what is published

• Which motorcycle dataset the 4800 N longitudinal stiffness was taken from
• Whether the paper's model couples longitudinal and lateral slip at all, or carries them independently
• The wheel loads carried by the front and rear wheels of the simulated bicycle, which fix the load at which the slip stiffness should be fitted
• The scatter in the Maier et al. measurements, and therefore the uncertainty on the fitted stiffness

This is the authors' own next step. They found that the load at which the tyre was measured moved the weave speed from 9.7 to 8.3 m/s, more than brand or pressure did, but the Magic Formula load scaling rests on just two measured loads, so the scaling cannot be checked and may be carrying the whole effect.

Question

Do Magic Formula coefficients fitted across a grid of vertical loads predict a different weave speed from coefficients scaled from the two loads of 400 and 600 N?

Hypothesis

Coefficients fitted over at least five vertical loads move the predicted weave speed by more than 0.2 m/s relative to the two-load scaling, and shrink its 1.4 m/s spread.

Attempt type

wet-lab
Vocabulary note: This is a bench measurement campaign on a tyre test rig followed by a simulation re-run. It is a physical experiment but nothing about it is wet or biological; wet-lab is simply the only physical-experiment value the vocabulary offers, and the honest label would be bench-measurement plus computational.

Design

One tyre specimen is measured on an existing bicycle tyre rig over a grid of at least five vertical loads bracketing the paper's 400 and 600 N and the 343 to 526 N range used by Dell'Orto et al. 2024, crossed with camber angles of minus 5, 0 and plus 5 degrees and inflation pressures spanning the paper's 2 to 5 bar. Lateral force, self-aligning torque and twisting torque are recorded at each condition. Two Magic Formula parameter sets are then fitted: one using only the 400 and 600 N conditions with the paper's load scaling, and one with load-dependent coefficients fitted across the whole grid, holding one load out as a test set. Both sets are fed into the full model and the measure is weave speed and the capsize re-stabilisation speed. Loads are added until the held-out weave-speed prediction moves by less than 0.2 m/s between refinements.

Controls

• known-positive-task Fit using only the 400 and 600 N conditions and check that the simulation pipeline reproduces the published weave speeds of 9.7 and 8.3 m/s.
  Rules out: That any difference between the two parameter sets is an artefact of a rebuilt fitting or simulation pipeline rather than of the load grid.
• within-subject One physical tyre specimen at one pressure setting across the whole load grid, with the pressure corrected for load-induced change before each run.
  Rules out: That a trend across loads is really a trend across specimens or a drift in inflation pressure.
• background The same reference condition, for example 400 N at 4 bar and zero camber, repeated at the start, middle and end of the campaign, plus a free-rolling run at each load.
  Rules out: That between-load differences are smaller than the rig's own repeatability or are caused by tread wear and warming over the session.
  Vocabulary note: Used as an instrument repeatability and drift floor, not a background signal.
• designated One load condition repeated on a second rig, for example the Padova rig that produced the original two-load dataset and the Politecnico di Milano VeTyT rig.
  Rules out: That the load dependence is a property of one rig's contact surface, drum curvature or temperature rather than of the tyre.
  Vocabulary note: In this field the designated control is a cross-instrument calibration repeat, not a designated positive control sample.

Sensitivity

The effect to be beaten is given in the record: the two-load scaling produces a 1.4 m/s spread in weave speed, 9.7 m/s from 400 N properties against 8.3 m/s from 600 N properties, and the smallest difference the paper treats as real is 0.2 m/s, so the design must resolve 0.2 m/s in weave speed. The measurement-side requirement, that is how precisely each coefficient must be measured for weave speed to be pinned to 0.2 m/s, cannot be derived from the record: the paper gives the coefficient sensitivities only as approximate percentage changes at 200 percent of nominal, around 25 percent for the twisting torque coefficient, 9 percent for camber stiffness and under 1 percent for cornering stiffness, and the rig repeatability is not stated anywhere. Deriving it needs the model's partial derivatives of weave speed with respect to each coefficient and a published repeatability figure for the rig.

Would count as

Supported → positive · opposed → negative-no-effect
Inconclusive if: The rig's repeatability on cornering stiffness or twisting torque turns out to be as large as the differences between loads, in which case the result is inconclusive-underpowered rather than a null, and the record gives no repeatability figure with which to predict this in advance.

Already done

• Lateral force and self-aligning torque of city and cargo bicycle tyres were measured under vertical loads from 343 to 526 N, camber angles of minus 5, 0 and plus 5 degrees and inflation pressures from 300 to 500 kPa, and Magic Formula parameters were identified; cornering stiffness generally falls with rising pressure for city tyres but rose for the cargo tyre between 300 and 400 kPa. (Dell'Orto, Mastinu, Happee and Moore 2024, Vehicle System Dynamics, doi:10.1080/00423114.2024.2338143)
  So: This is most of the authors' next step already done, so the design should be framed as extending the load range and feeding the parameters into the stability model rather than as a new measurement idea.
• A theoretical model relates bicycle tyre twisting torque to contact patch area and was validated against measurements with a mean error under 3 percent for camber angles below 10 degrees. (Dell'Orto, Ballo, Gobbi and Mastinu 2023, Measurement, doi:10.1016/j.measurement.2023.113460)
  So: Directly targets the coefficient this paper found dominant, and because contact patch area depends on load and pressure it supplies a physical form for the load dependence instead of an empirical scaling exponent.
• Internal pressure, vertical force, speed and rolling surface temperature all change the measured characteristics of a racing bicycle tyre, and higher inflation pressure raises cornering stiffness only at heavy vertical loads, while at low loads an over-inflated tyre has lower cornering stiffness. (Dell'Orto, Ballo, Mastinu, Gobbi and Magnani 2023, European Journal of Mechanics A/Solids, doi:10.1016/j.euromechsol.2023.105010)
  So: Shows load and pressure interact, so a one-dimensional load scaling applied at a single pressure is the wrong functional form and the grid must cross load with pressure.
• The sideslip and camber characteristics and mechanical properties of bicycle tyres were measured and identified for bicycle dynamics modelling, and this is the dataset of four tyres at 2 to 5 bar and 400 and 600 N that the paper used. (Doria, Taraborrelli and others 2011 and 2012, Vehicle System Dynamics, doi:10.1080/00423114.2011.615408 and doi:10.1080/00423114.2012.754048)
  So: Fixes the exact baseline the grid-fitted parameters must be compared against, and is the reason the two-load baseline cannot be refitted without the original raw data.
• Indoor measurements of the lateral characteristics of a cargo bicycle tyre extend bicycle tyre data to higher vertical loads than touring tyre campaigns. (Dell'Orto and others 2023, The Evolving Scholar, doi:10.24404/644ba1e13c57633bb23a9dc6)
  So: Shows the upper end of the load grid is already reachable with existing hardware, so the campaign does not need a new rig.

Searched: Searched OpenAlex for bicycle tyre cornering and camber stiffness against vertical load, load-dependent Magic Formula coefficients for bicycle tyres, bicycle tyre twisting torque, and bicycle tyre test rigs, filtered from 2015 onward, and followed the citing and related work of the paper's own Doria references. The Politecnico di Milano group has published four relevant campaigns since 2023; I did not find any paper that feeds load-dependent bicycle tyre coefficients back into an open-loop stability model and reports the resulting weave speed, which is the part of this next step that is still open.

Feasibility

A tyre measurement group with an existing bicycle tyre rig, such as Politecnico di Milano, Padova or Karlsruhe, paired with a multibody group to re-run the stability model · 6 to 12 months · moderate, because the rig and the fitting tools already exist and the real cost is rig time and one researcher
Blockers: The raw measurements behind the paper's two-load dataset are not public, so the two-load baseline has to be re-measured rather than refitted; The ADAMS model is not published, so the stability re-run needs a rebuilt model; The four specific tyres used in the paper may no longer be on sale, so the comparison may have to use a different specimen

Strongest objection

A specialist would say the load effect the paper reports is not a tyre property question at all but an extrapolation artefact: Magic Formula coefficients fitted at 400 N and then used at the actual static wheel loads of an 80 kg rider on a bicycle are being asked to work outside the range in which they were identified, and the cheap fix is to measure at the bicycle's own front and rear wheel loads rather than to build a load-scaling law. On that reading a six-load campaign is an expensive answer to a question that two well-chosen loads would settle.

Not derivable from what is published

• The rig repeatability on cornering stiffness, camber stiffness and twisting torque, which is needed to set the number of repeats per condition
• The front and rear static wheel loads of the Twade T3001 with the 80 kg, 1.80 m rider, which should set the centre of the load grid
• The functional form of the load scaling the paper used, and whether it scales all coefficients or only some
• Whether the 9.7 m/s figure belongs to case 6 or case 7, which the record flags as an unresolved inconsistency in the paper

This is the authors' own next step, and the record is blunt that nothing in the paper was checked against a real bicycle: validation here means agreement with published benchmark eigenvalues over 4 to 10 m/s. The authors want physical measurements to show that the known shortcomings of the Whipple model come from treating tyres as knife edges, and they also note that high speeds and the wobble mode were left out.

Question

Do measured weave and capsize eigenvalues of a real instrumented bicycle match the tyre-model predictions better than the knife-edge Whipple predictions across 4 to 12 m/s?

Hypothesis

Measured weave speed of a real bicycle lies closer to the tyre-model prediction of about 9.3 m/s than to the knife-edge prediction of 4.9 m/s, by more than the measurement uncertainty on weave speed.

Attempt type

wet-lab
Vocabulary note: This is a physical track experiment on a vehicle, paired with pre-registered simulation predictions. Neither wet-lab nor computational describes it; the vocabulary has no value for a field or track experiment and no value for a mixed physical and computational attempt.

Design

An instrumented riderless bicycle, fitted with sensors for roll rate, yaw rate, steer angle and rear wheel rotation as in Kooijman et al. 2007, plus frame-mounted accelerometers at several points to catch wobble as in Tomiati et al. 2020, is launched to coast freely on a level surface at a grid of speeds spanning 3 to 12 m/s. The same bicycle is run with two tyre sets: a measured set whose Magic Formula coefficients are known at that wheel load, and a hard narrow high-pressure set approaching the knife-edge idealisation. Its geometry and mass properties are measured so both models can be parameterised. Predictions from the knife-edge model and the tyre model are registered before any run. Weave and capsize eigenvalues are fitted from each free response, and the measure is the speed at which the weave real part crosses zero, with a confidence interval from repeated runs.

Controls

• known-positive-task Reproduce the Kooijman et al. 2007 result on the hard-tyre configuration first: measured eigenvalues that agree with the knife-edge benchmark in the low-speed low-frequency region.
  Rules out: That a mismatch with either model is caused by the launch rig, the sensors or the eigenvalue fitting rather than by the tyres.
  Vocabulary note: A published measurement standing in for a positive control.
• within-subject One bicycle frame, one sensor set and one launch procedure across both tyre configurations, with mass properties re-measured after the tyre swap.
  Rules out: That the difference between configurations comes from a different bicycle, different instrumentation or a different launch speed distribution.
• background Runs with the bicycle held upright and rolling with no lateral perturbation, and shaker or tap tests on the stationary frame, to record the noise floor and the frame and fork resonances.
  Rules out: That a peak identified as weave or wobble is sensor noise or a frame structural resonance, which the shimmy literature shows is a real risk at wobble frequencies.
  Vocabulary note: A sensor noise floor and structural resonance check rather than a biological background.
• designated A third configuration with solid non-pneumatic tyres, as the closest physical approach to a knife-edge wheel.
  Rules out: That the hard-tyre arm still has enough sideslip and camber compliance to be indistinguishable from the measured-tyre arm, which would make the comparison unable to separate the two models.
  Vocabulary note: A reference specimen bracketing the model idealisation, not a designated positive sample.

Sensitivity

The gap the design has to resolve is large and comes from the record: case 1 with stiff non-slipping tyres gives a weave speed of 4.9 m/s and case 6 with the Magic Formula tyre gives 9.3 m/s, a 4.4 m/s separation, which is more than twenty times the 0.2 m/s the paper treats as a small but real difference. What cannot be derived from the record is the run-to-run standard deviation of a measured weave eigenvalue, because the paper contains no physical measurements at all and the record does not carry Kooijman's scatter. Without that number the repeats per speed cannot be set, and the honest statement is that the number of runs must come from a pilot or from the published Kooijman et al. uncertainties.

Would count as

Supported → positive · opposed → refuted
Inconclusive if: Frame and fork compliance dominates the measured response, as the shimmy literature indicates it does at wobble frequencies, so that neither the knife-edge nor the tyre model fits and the mismatch cannot be attributed to tyre contact either way. A second route to inconclusive is that a riderless bicycle cannot be launched and kept upright above its stable range, leaving the high-speed capsize region unmeasured.

Already done

• An instrumented riderless bicycle coasting on a level surface was used to extract eigenvalues by curve fitting from measured roll rate, yaw rate and steer angle, and the knife-edge benchmark model proved fairly accurate for low-speed low-frequency behaviour. (Kooijman, Schwab and Meijaard 2007, Multibody System Dynamics, doi:10.1007/s11044-007-9050-x)
  So: This is the experiment the authors are asking for, already done for the knife-edge case, so the new content is the tyre comparison and the extension to higher speeds rather than the measurement itself.
• Linear accelerations and angular velocities at six points of a racing bicycle were recorded during a genuine hands-on shimmy; the rear frame rotates rigidly about a yaw-parallel axis while undergoing relative torsion between head tube and dropouts, and the front frame rotates about the steering axis while bending laterally. (Tomiati, Magnani and Marcon 2020, Vehicle System Dynamics, doi:10.1080/00423114.2020.1762902)
  So: Shows that the wobble half of this issue is governed by frame compliance, which this paper's model does not include, so instrumentation must separate frame modes from tyre effects or the result will be uninterpretable.
• The wobble mode of a bicycle and the wobble of a racing bicycle with the rider's hands on and off the handlebar were analysed and measured. (Magnani and others 2011 and Klinger and others 2014, Vehicle System Dynamics, doi:10.1080/00423114.2011.594164 and doi:10.1080/00423114.2013.877592)
  So: Both are already in this paper's reference list and the Klinger comparison value of about 9.5 m/s is used in the record, so they set the comparator for the wobble extension.
• Stability indices from eigenvalues of weave, capsize and wobble were computed with analytical bicycle models of increasing complexity including frame compliance, tyre properties and rider posture, and once tyre sliding is included the capsize mode is always stable, so the simpler models mispredict global stability even where they get single modes right. (Doria, Roa and Munoz 2019, Mechanical Sciences, doi:10.5194/ms-10-229-2019)
  So: Independently supports the paper's capsize claim by a different route and narrows what the measurement needs to settle to the weave speed and the high-speed region.
• Experiments on an autonomous bicycle were used to test and then modify a linear model and control law for steering and stability. (Xiong, Yu and Liu 2024, Nonlinear Dynamics, doi:10.1007/s11071-023-09249-1)
  So: Offers an alternative platform in which a controlled bicycle can be driven into speeds a free-coasting riderless bicycle cannot reach, at the price of having a controller in the loop.
• The Whipple and Carvallo stability model was applied to seven measured real bicycle configurations with and without pedal assistance, by analysis only, with no physical dynamic measurement. (Dieltiens, Debrouwere, Juwet and Demeester 2020, Applied Sciences, doi:10.3390/app10165672)
  So: Confirms that eleven years on, the usual practice is still to apply the model to measured geometry rather than to validate its eigenvalues against measured dynamics, so the gap the authors named is still open.

Searched: Searched OpenAlex for experimental validation of the Whipple bicycle model, instrumented bicycle lateral dynamics measurement, riderless bicycle measured eigenvalues, and bicycle wobble and shimmy high-speed stability, mostly filtered from 2015 onward, and resolved the paper's own reference list. I also checked the first author's 2017 Twente thesis: its abstract describes measured bicycle and rider interaction forces, kinematics and an identified rider balance controller, not measured open-loop weave and capsize eigenvalues, so it does not appear to close this issue. I found no post-2015 study that measures open-loop weave eigenvalues on a real bicycle and compares them against both a knife-edge and a measured-tyre model.

Feasibility

A bicycle dynamics group with an instrumented bicycle, a launch method and a large flat test surface, for example Delft, Twente, Padova or the Davis group, with tyre characterisation bought in from a rig group · 9 to 18 months · moderate to high, because it needs instrumentation, a large smooth test surface, tyre rig time for the same specimens, and safety provision for repeated riderless launches
Blockers: Launching a riderless bicycle repeatably and safely above about 6 m/s; The capsize-unstable high-speed region cannot be sustained for long enough to fit, which is the same limitation as iss-1 in physical form; The paper's ADAMS model is not published, so both predictions have to be rebuilt before they can be registered; No hard tyre is a true knife edge, so the comparator arm is an approximation of unknown quality

Strongest objection

A specialist would object that a riderless bicycle cannot validate this paper at all. The paper's conclusions are about a bicycle plus a passive rider of 80 kg, whose arm stiffness alone can destroy low-speed stability, and that mass dominates the system; remove the rider and you are measuring a different vehicle. Put the rider back and you cannot switch off their control, so an open-loop measurement is unobtainable in principle and the comparison has to be made against an identified rider controller instead. The second objection is that any mismatch will be attributed to frame and fork compliance before it is attributed to tyre contact, because the shimmy measurements show compliance is already in play.

Not derivable from what is published

• The run-to-run standard deviation of a measured weave eigenvalue, and therefore the required number of runs per speed
• The geometry and mass parameter set of the Twade T3001 bicycle used in the paper
• Whether the capsize mode is observable at all in a free coasting response
• A launch method that reaches 12 m/s riderless
• The Magic Formula coefficients of the specific tyres fitted to whichever bicycle is instrumented