aicognition/mrsd-exo-ankle

MRSD Ankle Exoskeleton — Torque Prediction Dataset

Synchronised wearable-sensor and inverse-dynamics data for training models that predict the human ankle plantarflexion moment from IMUs, an ankle angle and a ground-contact flag. Used to drive assistive torque on the MRSD ankle exoskeleton.

• Subjects: 22 (AB06–AB30), able-bodied
• Task: treadmill walking, 0.86–1.29 m/s
• Trials: 157 (~7 per subject, ~140 s each)
• Rates: IMU/ID 200 Hz, force plate/goniometer 1000 Hz, gait phase 200 Hz
• Source: derived from Camargo et al. 2021 (Georgia Tech EPIC lab); reshaped to Parquet, unused modalities (EMG, markers, IK, joint power) dropped

Repository layout

metadata.parquet                              one row per (subject, trial)
subjects/<subject>/<trial>__imu.parquet        foot/shank/thigh/trunk accel+gyro
subjects/<subject>/<trial>__id.parquet         inverse-dynamics joint moments (N·m)
subjects/<subject>/<trial>__fp.parquet         treadmill force plate (GRF, CoP)
subjects/<subject>/<trial>__gon.parquet        goniometer joint angles (degrees)
subjects/<subject>/<trial>__gcRight.parquet    right-leg gait phase (HeelStrike, ToeOff)

Every sensor file has a time_s column (seconds) plus that sensor's channels, at native sample rate — no resampling or filtering is applied at rest.

metadata.parquet columns

subject, trial, terrain, n_imu_samples, age, gender, height_m, weight_kg, speed_mean_mps, trialstarts_s, trialends_s.

Sensor channels

File	Rate (Hz)	Key channels	Units
imu	200	{foot,shank,thigh,trunk}_{Accel,Gyro}_{X,Y,Z}	m/s², rad/s
id	200	ankle_angle_r_moment (target), plus hip/knee/pelvis/lumbar moments	N·m
fp	1000	Treadmill_{R,L}_{vx,vy,vz,px,py,pz,moment_x,moment_y,moment_z}	N, m, N·m
gon	1000	{ankle,knee,hip}_{sagittal,frontal}	degrees
gcRight	200	HeelStrike, ToeOff	percent of cycle (0–100), resets at the event

Prediction target

id.ankle_angle_r_moment — the right-ankle net plantarflexion/dorsiflexion moment from OpenSim inverse dynamics, in N·m. Plantarflexion (push-off) is negative in this convention. For subject-independent reporting, divide by metadata.weight_kg to get N·m/kg.

NaN in the target marks samples outside the valid inverse-dynamics window.

Download

pip install huggingface_hub pandas pyarrow

Whole dataset:

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="aicognition/mrsd-exo-ankle",
    repo_type="dataset",
    revision="v1.0",
    local_dir="mrsd-exo-ankle",
)

One subject only:

root = snapshot_download(
    repo_id="aicognition/mrsd-exo-ankle", repo_type="dataset", revision="v1.0",
    allow_patterns=["metadata.parquet", "subjects/AB06/*"],
    local_dir="mrsd-exo-ankle",
)

Load a trial

import pandas as pd
from pathlib import Path

root = Path("mrsd-exo-ankle")
meta = pd.read_parquet(root / "metadata.parquet")
subject, trial = "AB06", "treadmill_01_01"

imu = pd.read_parquet(root / f"subjects/{subject}/{trial}__imu.parquet")
idf = pd.read_parquet(root / f"subjects/{subject}/{trial}__id.parquet")
gon = pd.read_parquet(root / f"subjects/{subject}/{trial}__gon.parquet")
gc  = pd.read_parquet(root / f"subjects/{subject}/{trial}__gcRight.parquet")

mass = meta.loc[(meta.subject == subject) & (meta.trial == trial), "weight_kg"].item()
ankle_moment_nm_per_kg = idf["ankle_angle_r_moment"] / mass

Visualise one gait cycle

import numpy as np
import matplotlib.pyplot as plt

# heel-strike events = where the gait-phase signal wraps back to 0
hs = np.where(np.diff(gc["HeelStrike"].to_numpy()) < -50)[0] + 1
tau = (idf["ankle_angle_r_moment"] / mass).to_numpy()

cycles = []
for a, b in zip(hs[:-1], hs[1:]):
    seg = tau[a:b]
    if 60 < len(seg) < 400:
        cycles.append(np.interp(np.linspace(0, 1, 100),
                                np.linspace(0, 1, len(seg)), seg))
cycles = np.array(cycles)

phase = np.linspace(0, 100, 100)
plt.fill_between(phase, cycles.mean(0) - cycles.std(0),
                 cycles.mean(0) + cycles.std(0), alpha=0.2)
plt.plot(phase, cycles.mean(0), lw=2)
plt.gca().invert_yaxis()                       # plantarflexion is negative
plt.xlabel("gait cycle (%)"); plt.ylabel("ankle moment (N·m/kg)")
plt.title(f"{subject} — mean ± SD over {len(cycles)} strides")
plt.show()

The mean-cycle push-off peak is around −1.5 to −2 N·m/kg (single strides reach −1.8 to −2.5), near zero through swing.

Suggested split (subject-independent)

Split	Subjects
train	AB06, AB07, AB08, AB11, AB12, AB13, AB14, AB15, AB16, AB17, AB18, AB19, AB20
val	AB09
test	AB10, AB21, AB23, AB24, AB25, AB27, AB28, AB30

Fit any normalisation on train subjects only.

Full training pipeline

The mrsd_exo codebase ingests this dataset (100 Hz resample with anti-aliasing, binary stance from the gait-phase signal, subject-wise scaling) and trains a causal TCN. See the project repository for scripts/prepare.py and scripts/train.py.

Source, changes made, and licence

This is a derived work of the Georgia Tech / EPIC-Lab lower-limb biomechanics dataset by Camargo et al. (2021), used and redistributed under the Creative Commons Attribution 4.0 International licence (CC BY 4.0).

• Original dataset: Camargo, Jonathan (2021), "A comprehensive, open-source dataset of lower limb biomechanics ... Part 1 of 3", Mendeley Data, V2, DOI 10.17632/fcgm3chfff.2. Download and mirror: https://www.epic.gatech.edu/opensource-biomechanics-camargo-et-al/
• Original publication: Camargo, J., Ramanathan, A., Flanagan, W., Young, A. (2021). A comprehensive, open-source dataset of lower limb biomechanics in multiple conditions of stairs, ramps, and level-ground ambulation and transitions. Journal of Biomechanics, 119, 110320, DOI 10.1016/j.jbiomech.2021.110320.
• Licence: CC BY 4.0 — https://creativecommons.org/licenses/by/4.0/legalcode. This derived repository is redistributed under the same licence.

Changes made in this derived version:

1. Kept only the treadmill ambulation condition (stairs, ramps, level-ground overground, and walk-to-run transitions were excluded).
2. Kept only the imu, id, fp, gon, and gcRight sensor streams; excluded emg, markers, ik, ik_offset, jp (joint power), and gcLeft.
3. Reorganised from the original per-sensor CSV directory tree into one Parquet file per (subject, trial, sensor), with the original Header column renamed to time_s. Signal values, sample rates, and units are unchanged.
4. Consolidated per-subject demographics and per-trial condition files (subject_info, speed, trialStarts, trialEnds) into a single metadata.parquet.

No signal data was resampled, filtered, or otherwise altered in value; the resampling and feature engineering used for model training happen downstream in the mrsd_exo codebase, not in this dataset.

No endorsement. Nothing here implies that Georgia Tech, the EPIC Lab, or the original authors endorse this derived dataset or its use.

Third-party content. Any material within the original dataset identified as belonging to a third party may require separate permission; this redistribution grants no additional rights over such material.

Citation

Camargo, J., Ramanathan, A., Flanagan, W., Young, A. (2021). A comprehensive, open-source dataset of lower limb biomechanics in multiple conditions of stairs, ramps, and level-ground ambulation and transitions. Journal of Biomechanics, 119, 110320.