Data and code associated with the paper: "A velocity-loop view of avian gait: hodograph rotation sense characterizes the walking-to-running continuum in guinea fowl" Authors: Monica A. Daley and Aleksandra Birn-Jeffery
Biology Open (2026). doi:10.1242/bio.062880

*Author for correspondence: madaley@uci.edu

Version

This version corrects the sign of the horizontal ground reaction force in 77 trials and every quantity computed from it. Version History, below, lists the changes and their effect on the results the article reports, which were computed from the previous version.

Overview

This package contains ground reaction force recordings and reconstructed center-of-mass (CoM) trajectories for guinea fowl (Numida meleagris) running over level ground, compiled from three Royal Veterinary College data collections and reprocessed through one common pipeline. The analysis asks how the rotation sense of the CoM velocity loop, the hodograph, relates to the transition from pendular (walking) to bouncing (running) dynamics, and compares it against established center-of-mass criteria across a speed range that includes extensive grounded running.

Each of 334 trials is provided as a whole-recording time series in a standardized, bird-centered coordinate frame, together with tidy per-step and per-stride measure tables, per-step gait classification, and ensemble mean step cycles by gait and speed.

Directory Structure

GaitSel_DryadPackage_AllGF/
├── README.md                                      # This file
├── trial_index.csv                                # One row per trial (334)
├── steps_index.csv                                # One row per step (3842)
├── perStep_long_multi.csv                         # Tidy per-step measures (3842 x 90)
├── perStride_long_multi.csv                       # Tidy per-stride measures (1678 x 78)
├── morphology_multi.csv                           # Per-session body mass and leg length (30)
├── trialRoster_multi.csv                          # Source file provenance per trial (334)
├── force_sign_corrections.csv                     # Force columns whose sign this version corrects (78)
├── per_trial_timeseries.zip                       # Whole-bout time series, one pair per trial
│   └── per_trial_timeseries/                      #   unzips to this folder
│       ├── [bird]_[YYYYMMDD]_[trialStem].csv      # 20 channels at 500 Hz
│       └── [bird]_[YYYYMMDD]_[trialStem].mat      # Same, plus metadata and a step table
└── mean_stepcycle_by_gait_speed.zip               # Ensemble mean step cycles
    └── mean_stepcycle_by_gait_speed/              #   unzips to this folder
        ├── mean_cycles.csv                        # Long format, 10800 rows
        └── mean_cycles.mat                        # Same content as a MATLAB struct

The two folders are deposited as .zip archives because Dryad stores a flat file list and would
otherwise drop the folder they belong to. Each archive unzips to the folder named above, so the
layout after unzipping is the one the pipeline and the paths below expect. Everything else is a
flat file at the top level.

Data Files

File Naming Convention

Per-trial files follow the naming pattern:

[bird]_[Date]_[trialStem].csv

Example: rvc_bly_20090505_L00cm_0009.csv

The same string is the boutID, which is the join key across every table in this package.

Studies Included

All level (unperturbed) running trials with raw force and synchronized marker capture.

Force was sampled at 500 Hz.

Subject Information

Thirty bird-sessions over thirteen individuals. Colour codes name the same individual within a collection cohort but were reused across cohorts, so subjectID is the reconciled identity and is the correct grouping factor for any analysis pooling across studies. The two 2012 Blum studies, four months apart, are one cohort. The June bird recorded as noc, carrying no colour band, is the blue bird: the study's weight record lists blue among the seven birds run and the archive holds no blu for that study, it is the only unmatched bird, and the original analysis's published data uses blue's weighed mass for 'noc'. Body mass and leg length are per session, since a bird's mass differs between sessions.

Leg length L0 is the median CoM-to-foot distance at touchdown over the session's steps with no aerial phase. Where a session yielded fewer than five such steps it falls back to that bird's sessions pooled, and for the one bird that never yielded five, to a within-sample allometry on body mass; the L0 source column records which applies.

Data Selection

Every detected step and stride is provided, with flags rather than deletion, so a reader can reproduce the paper's sample or define a different one.

3842 steps detected
2618 pass the quality-control gate           (qcPass = 1)
   0 duty factor undetermined
  40 touchdown height out of range
   0 both
2578 in the analysis sample                  (analysisSample = 1)

Two gait-space descriptors depend on foot-marker quality that cannot be detected based on a force-based gate: the per-limb duty factor is undefined where foot contact could not be determined from the markers, and the touchdown energy is considered infeasible where the reconstructed CoM height above the contacting foot falls outside 0.4 to 1.6 leg lengths.

analysisSample = 1 is the column to filter on to reproduce the paper. It selects 2578 steps and 942 strides. Within the step sample the steadiness split is 929 steady, 942 accelerating and 707 decelerating. Of 1678 strides detected, 1126 pass the stride gate and 942 are in the analysis sample, those being the strides that pass the gate and whose two steps are both in the step sample.

---

Missing Values

Every CSV in this package writes a missing value as NA, and no cell is left blank. A value is
NA where the quantity is not defined for that row, and the flag or label that says why it is
undefined sits in the same row. Nothing is missing through a lost or corrupted recording.

Every other column has a value in every row. dutyFactor, the marker-based per-limb duty factor
the paper reports, is defined for all 3842 steps, as is the dimensionless speed u. The 334
per-trial time-series CSVs and mean_cycles.csv contain no missing values at all.

The .mat files hold the same content in MATLAB types. In the steps table
inside each per-trial .mat, the three label columns are empty strings ("") on the 1222 rows
where the CSV reads NA, which is how MATLAB represents an unset string, and numeric quantities
use NaN.

---

per_trial_timeseries Folder

Whole on-plate bout per trial on the 500 Hz force time base, as a CSV and a MATLAB .mat pair.
All values are in SI units and in the bird-centered frame described under Coordinate Frame below.

Trajectory File Columns (.csv)

The reconstructed CoM and the proxy are both provided so a reader can see the reconstruction quality directly for any trial.

MAT File Contents

Each .mat holds one trial struct with the same channels as the CSV plus:

• Acquisition metadata: boutID, bird, subjectID, study, dateCode, colourCode, mass_kg, L0_m, bodyWeight_N, fs_force_Hz, fs_kin_Hz, g, nPlates, foreaftAxis, axisOrder
• Reconstruction quality: driftRMS_vert_mm, WE_identity_r
• steps: a table of this trial's rows of steps_index.csv, with the same 14 columns

---

trial_index.csv

One row per trial (334 rows, 15 columns).

---

steps_index.csv

One row per step (3842 rows, 14 columns). The compact step-level index; the full measures are in perStep_long_multi.csv.

---

perStep_long_multi.csv

The tidy per-step analysis table: 3842 rows, 90 columns, one row per step. This and perStride_long_multi.csv are the tables the analysis and the summary statistics are computed from.

Columns ending _n are dimensionless (see Normalisation). Raw and dimensionless forms are both kept so a reader can work in either.

Identity and grouping

Classification and sample flags

Timing and duty

Speed and distance

Forces and impulse

Center-of-mass energy and work

Gait descriptors (the six continuous descriptors of the gait space, plus the hodograph measure)

Virtual leg and leg stiffness

legCompress_n and kLeg_n are NA where the inversion returns a non-physical compression. Leg stiffness is defined only relative to the spring-mass template it comes from, and where two limbs share the net force in double support it is an effective two-limb stiffness rather than a single-limb property.

---

perStride_long_multi.csv

The tidy per-stride table: 1678 rows, 78 columns, one row per stride (two steps). Column meanings match perStep_long_multi.csv except as noted.

Pendular recovery is R = (W_f + W_v - W_tot)/(W_f + W_v), each W the sum of positive increments over the stride, and is defined per stride because the CoM returns to its state over a full stride.

---

morphology_multi.csv

One row per study session (30 rows, 10 columns).

---

trialRoster_multi.csv

Source-file provenance, one row per trial (334 rows, 12 columns): boutID, bird, colourCode, study, dateCode, trialNo, trialStem, bodyMass_kg, windowMode, and the forceFile, markerFile and logFile paths of the raw recordings.

Paths are relative to the root of the source archive they were collated in, for example
OtherSpecies_byDate/Guinea fowl/Blum_Surface_2012-06/2012-06-25/gf_red_LE0cm_0006.txt, so they carry no machine or user name. logFile is NA for the trials that have none: only the 68 Blum_Surface_2012-06 trials were recorded with one.

The raw force and marker files are not part of this deposit and are not needed to reproduce the analysis, which runs from per_trial_timeseries/. This table is provenance: it records which source recording each trial came from, so a trial here can be traced back to the collection it belongs to.

---

force_sign_corrections.csv

One row per corrected force column (boutID, axis, column, reason): the trials and the axis, ml or fa, whose ground reaction force has the opposite sign in the previous version. Version History describes the correction.

---

mean_stepcycle_by_gait_speed

Ensemble mean step cycles for each gait by speed-bin group, in long format (10800 rows) and as a MATLAB struct. Six groups: walk, grounded run and aerial run, each split slow and fast.

Two things determine which cycles enter a group, which corresponds to the paper's mean-trace figure. 1) The cycles are the STEADY steps of the analysis sample (steadiness = steady and analysisSample = 1). 2) The slow/fast boundary is the gait's median Froude number over all of its classified steps, not the median of the steady subset, so the boundary does not move between steadiness classes; the two halves of a group can therefore be unequal in size.

Channels: F_vt_N, F_fa_N, F_ml_N, F_vt_BW, F_fa_BW (forces, raw and body-weight normalised); vfa_ms, vvert_ms (CoM velocity); com_fa_disp_m, com_vt_rel_m (CoM displacement); KE_J, PE_J, Etot_J, KE_n, PE_n (energy, raw and normalised); stanceFoot_fa_relCoM_m, stanceFoot_vt_relCoM_m, swingFoot_fa_relCoM_m, swingFoot_vt_relCoM_m (foot position relative to the CoM).

---

Coordinate Frame and Units

Every three-vector is ordered [medio-lateral, fore-aft, vertical]. The frame is bird-centered and right-handed: fore-aft is positive in the direction of travel for every trial, vertical is positive up, and medio-lateral completes the right-handed frame. Trials in which the bird crossed the runway in the opposite lab direction have their fore-aft and medio-lateral signs flipped together, so all trials are directly comparable without further reorientation.

Units are SI: force in newtons, position in meters, velocity in meters per second, time in seconds, mass in kilograms, energy in joules. Angles are in degrees except the collision angle, which is in radians because the equivalence with the dimensionless cost of transport holds in that form.

Normalisation

Data are normalised into dimensionless quantities following dynamic similarity, using each session's body mass m, leg length L₀ and g = 9.81 m/s². Columns ending _n carry the dimensionless form.

Dimensionless speed is u = v/√(gL₀) and the Froude number is u².

---

Processing Workflow

The analysis code is on GitHub at https://github.com/NeuroMechLab/GuineaFowlGait-code, tagged v1.3.0. PIPELINE.md there gives the full run order and maps every output to the manuscript element that uses it.

Reproducing the analysis from this package

The release runs from this package. The per-trial files carry the net ground reaction force, the time base, the kinematic CoM proxy the reconstruction is path-matched to, both foot marker tracks, and body mass, leg length and g, which is everything the measurement kernels read, so the raw force-plate and marker recordings are not needed.

Unzip per_trial_timeseries.zip and mean_stepcycle_by_gait_speed.zip inside the package, put the package at GaitSel_DryadPackage_AllGF/ beside _MatlabProcessingCode/, then:

matlab -batch "addpath(genpath('_MatlabProcessingCode')); \
    GaitSelMulti_BatchFromDryad; GaitSelMulti_ExportTidyCSV; \
    GaitSelMulti_QCGate; GaitSelMulti_ExportCycleTraces"
cd _RAnalysis && Rscript run_all.R

Step 1. GaitSelMulti_BatchFromDryad.m rebuilds each trial from per_trial_timeseries/ and re-runs the reconstruction (GaitSel_ReconstructCoM), the gait-event detection at foot-marker touchdowns (GaitSel_DetectGaitEvents with GaitSel_FootContacts) and the per-step and per-stride reduction (GaitSel_PerStepStrideMeasures).

Step 2. GaitSelMulti_ExportTidyCSV.m derives per-session L0, adds the dimensionless and leg-stiffness columns and flags outliers, writing perStep_long.csv, perStride_long.csv and morphology.csv.

Step 3. GaitSelMulti_QCGate.m applies the work-energy and drift criteria and the energy-grade steadiness criterion. This is the single definition of the retained set, computed once in MATLAB so that this package and the statistics agree.

Step 4. GaitSelMulti_ExportCycleTraces.m writes the per-step cycle traces the figures read, also from per_trial_timeseries/.

Step 5. Rscript run_all.R runs the numbered R chain: 01_load.R reads the tidy tables and builds the individual grouping, 02_clean.R applies the quality-control and steadiness flags, 03_step_descriptors.R adds the collision angle and touchdown energy, 04_analysis_sample.R fixes the analysis sample of 2580 steps and 944 strides, and the remaining scripts produce every figure, table and statistic. GaitSel_ExportDryadMeanCycles.m rebuilds mean_stepcycle_by_gait_speed/ afterwards, since it reads the R phase's speed-bin thresholds.

Nothing is written back into this package: the re-run's outputs go to rerun/ and SavedResults/, so a second run starts from the archived state. rerun/reconstruction_check.csv compares the re-run against the values these files already carry. Over all 334 trials it agrees to 0.000 mm on the vertical CoM path and on the drift, to 0 on the work-energy identity ratio, and on the number of steps cut for every trial.

One column cannot be reproduced from this package. dutyLimb, the plate-based duty-factor cross-check, is estimated from the per-plate vertical force, which is not deposited, so it comes back as NA. The duty factor the paper reports is the marker-based dutyFactor and is unaffected.

How this package was made

The per-trial files themselves come from the raw force-plate and marker recordings, which are not part of this deposit. Those were paired into a trial roster, imported and windowed, the horizontal force axes resolved to anatomical fore-aft and medio-lateral, the CoM offset applied to the marker proxy and the foot tracks de-spiked, before the reconstruction and measurement steps above. That import layer has no input a reader can supply and is not part of the code release.

Reconstruction method

The CoM is reconstructed by path-matched double integration of the ground reaction force over the whole on-plate bout: F = ma, with one set of integration constants per axis (initial position, initial velocity and an acceleration-baseline offset) fitted to minimise departure from the kinematic CoM proxy. The net force is low-pass filtered at 50 Hz with a zero-phase Butterworth filter. The analysis window for each trial is found automatically as the span where the bird is both tracked and on the plates. The orientation of the force plates relative to the motion-capture frame is fixed within a collection date and differs between collection years, so the horizontal force axes are assigned once per date. Fore-aft is the plate axis, with the sign, whose cumulative impulse best follows the kinematic fore-aft velocity over all of that date's trials. Medio-lateral is the other plate axis, with the sign that keeps the plate frame and the motion-capture frame both right-handed.

The proxy is the midpoint of the cranial and caudal back markers, which lies above and behind the whole-body CoM, so a fore-aft and a vertical offset are applied to it before any geometry is taken. The fore-aft offset is fitted by driving the net pitch impulse about the CoM to zero over stride-complete windows, within anatomical bounds set by the two back markers. That condition does not identify the vertical offset, because the fore-aft impulse of the summed force approaches zero over a whole stride, so the vertical offset is set from anatomical scaling as -0.2625 L_iso with L_iso = 0.20 m^(1/3). Offsets are fitted for each bird in each recording session, because marker placement was renewed at each session. comProxy_* in the per-trial series carries them, so leg length, leg angle and CoM height above the foot all refer to one CoM. Energy fluctuations are unaffected, because a constant shift cancels in a height change.

Reconstruction quality per study, as median work-energy identity r and median vertical drift: RVC 0.99 and 7.7 mm, DropVsPothole 1.00 and 7.5 mm, Surface 0.97 and 4.9 mm.

---

Version History

This version: horizontal force sign correction

This version corrects two errors in the sign of the horizontal ground reaction force (F_ml_N, F_fa_N; GRF_N in the .mat files).

• Medio-lateral force, in all 74 trials recorded on 2008-11-20, 2008-11-21, 2009-11-17, 2009-11-18, 2009-11-19. In these sessions the force plates were turned 180 degrees relative to the motion-capture frame, and the medio-lateral axis had been given the orientation of the other sessions.
• Fore-aft force, in 4 trials whose own sign fit disagreed with the other trials of their collection date: rvc_red_20081120_L00cm_0006, surf_noc_20120626_LE0cm_0003, surf_red_20120625_LE0cm_0004, surf_yel_20120627_LE0cm_0009. rvc_red_20081120_L00cm_0006 is one of the four rvc_red trials of 2008-11-20 whose force and marker files come from different recordings; its sign follows the other trials of that date, and none of those four trials has a step in the analysis sample.

Integrated over body mass, the listed force components changed the CoM velocity in the direction opposite to the velocity of the CoM proxy, the back-marker midpoint. With the corrected signs, the reconstructed CoM follows the proxy: over the corrected trials, the median root-mean-square distance between them in the medio-lateral direction is 7.5 mm, against 27.6 mm in the previous version.

Recomputed from the corrected force: the CoM path and velocity, the drift and the work-energy identity of the corrected trials; the leg length L0 of the 12 sessions in which it changes by more than one micrometre, since L0 is a median CoM-to-foot distance (the largest changes are rvc_yel|2008-11, 0.195 to 0.163 m, and rvc_blu|2008-11, 0.203 to 0.187 m); every quantity normalised by L0 in those sessions; the per-step and per-stride measures, labels and flags of the 95 affected trials; and the mean step cycles. 2 of the affected trials lie outside the corrected sessions: the mis-cut-cycle filter compares each stride with strides of similar dimensionless speed in all trials, and two of their flags change. Every other row and per-trial file is identical to the previous version.

In the November 2008 sessions of rvc_yel and rvc_blu, the corrected leg lengths are shorter than in the same birds' later sessions (rvc_yel 0.163 m against 0.197 m in April 2009, rvc_blu 0.187 m against 0.200 m in May 2009), because the height of the CoM above the stance foot at touchdown is lower in that collection (0.134 m against 0.180 m for rvc_yel), which is likely due to differences in back-marker placement between recording sessions; these two sessions provide 30 of the 2578 analysis-sample steps, so the difference is unlikely to shift any major result of the article.

The analysis sample is 2578 steps and 942 strides, against 2580 and 944. Of the 89 values the article reports, 51 change. The three crossover speeds are u = 1.057 (Froude 1.116), u = 0.560 (Froude 0.313) and u = 0.576 (Froude 0.332) (article: u = 1.056 (Froude 1.114), u = 0.557 (Froude 0.310) and u = 0.576 (Froude 0.332)); the agreement between the two-cluster split and rotation sense is 87.7% (2260 of 2578 steps) (article: 87.7% (2262 of 2580 steps)); and the leg-length range is 0.163-0.232 m (article: 0.192-0.232 m).

Although the numbers shift slightly, none of the claims in the paper are changed by this update. The main claims hold at the precision reported in the paper. Crossover speeds: rotation sense u = 0.58 (0.46 to 0.63) and footfall u = 1.06 (1.01 to 1.09) are unchanged when rounded as in the text. The energy-phase crossover stays at u = 0.56. The rotation-sense and energy-phase intervals still overlap, and the footfall transition is still roughly twice as fast. Agreement with the cluster split: 87.7% unchanged. Descriptive ordering across walk, grounded run and aerial run: each median moves by at most 0.4 units, and the order is the same. Leg stiffness: 7.1, 6.2 and 9.7 for those three gait categories is unchanged.

---

Code Dependencies

MATLAB R2019b or later, with the Statistics and Machine Learning, Signal Processing and Curve Fitting Toolboxes.

R 4.6.0, with dplyr, ggplot2, patchwork, readr, tidyr, mgcv, cluster, openxlsx and knitr.

All scripts locate the project root from their own file location, so no path configuration is required.

---

License

Data in this deposit are released under CC0 1.0 Universal public domain dedication, as Dryad requires: https://creativecommons.org/publicdomain/zero/1.0/

The analysis code on GitHub is released separately under the MIT License.

---

Citation

Please cite the associated manuscript when using this data or code:
Daley, M. A. and Birn-Jeffery, A. (2026). A velocity-loop view of avian gait: hodograph rotation
sense characterizes the walking-to-running continuum in guinea fowl. Biology Open,
doi:10.1242/bio.062880

Please also cite this data package:

Daley, M. A. and Birn-Jeffery, A. (2026). Data from: A velocity-loop view of avian gait:
hodograph rotation sense characterizes the walking-to-running continuum in guinea fowl. Dryad,
doi:10.5061/dryad.7m0cfxqc4

Contact: For questions regarding this dataset or code, please contact the corresponding author Monica Daley (madaley@uci.edu).