amathislab/Boo_LevelWalking

Policy Reference NPZ Data

MuscleMimic paper: Towards Embodied AI with MuscleMimic: Unlocking full-body musculoskeletal motor learning at scale.

This dataset contains per-subject, per-trial .npz files exported from OpenSim inverse dynamics results for level walking. It also has muscle activations and forces from static optimization.

The files are intended for comparing policy inference outputs against OpenSim ID torques. Each exported trial uses the kinematics timeline as the master timeline, with time and timestep reset to start from zero.

ID torques are normalized by the subject's weight in kg from Subject_Demographics.xlsx. GRF force components are normalized by body weight in Newtons, computed as subject_weight_kg * 9.80665.

Source data: Gait120 (Koo, Boo, Seo, Kim, 2025), doi:10.6084/m9.figshare.27677016, CC BY 4.0.

Coverage: 419 trials from 101 subjects (S011–S120). This is every trial with both Step01 and Step02. S001–S010 have no force-plate data in Gait120, so they are not included.

Processing Notes

• Model. The generic OpenSim Rajagopal2016.osim is scaled per subject. Gait120 has no static trial, so segment scale factors come from median marker distances over all walking frames. Hip joint centres use the Harrington regression. Mass comes from Subject_Demographics.xlsx.

• Windows. Step01 and Step02 kinematics and GRF are combined into one Step01_Step02 window.

• Sign conventions. The Gait120 joint angles come from a modified Rajagopal model with two differences from Rajagopal2016:

  • knee extension is positive;
  • the left hip adduction and rotation axes are mirrored.

  Both are converted before ID. This was checked against the markers: the left leg sits 8–17 cm from its markers with the values as given, and ~2 cm after the correction. Every joint_torque column is reported in the same sign convention as its joint_angle column.

• Locked coordinates. Coordinates locked in Rajagopal2016 (subtalar, mtp, wrist) are held at 0 during ID. joint_angle keeps the raw Gait120 values.

• Filtering. Kinematics and GRF are both low-pass filtered at 6 Hz. GRF force and moment are filtered, and the COP is then rebuilt.

• Force plates. Each plate contact is assigned to calcn_r or calcn_l by its COP distance to the TRC heel/toe markers. A contact is rejected (both feet on one plate, or a partial hit) when its mean COP is more than 10 cm from the assigned foot.

• Validity. Gait120 has only two force plates, so a foot often lands off them. Foot contact is detected from heel/toe marker speed. joint_torque is NaN wherever it is not valid, judged per column:

  • leg columns are valid when that foot is in swing or fully on a plate;
  • pelvis columns are valid only when every grounded foot is measured;
  • lumbar columns are always valid.

  A leg's joint moments depend only on that leg's segments and its own GRF, so one leg can be valid while the other foot is down off-plate.

• QC. The median RMS pelvis residual in fully measured frames is 2.3 % BW (ML), 5.4 % BW (vertical) and 3.2 % BW (AP). Per-trial flags are in the qc_notes key. Two are worth knowing:

  • S036's GRF suggests the recorded body mass is too low, so its per-kg values are inflated.
  • S100 Trial02 has a partial plate hit on the left foot.

The exported computed_grf is the filtered GRF used for inverse dynamics, in per-plate columns. Force-vector columns are normalized by body weight in Newtons in the .npz; COP and ground-torque columns remain in physical units.

Folder Structure

S011/
  S011_Trial01_Step01_Step02.npz
  S011_Trial01_Step01_Step02_muscle_activation.sto
  S011_Trial01_Step01_Step02_muscle_force.sto
  S011_Trial01_Step01_Step02_muscle_activations.png
  S011_Trial02_Step01_Step02.npz
  ...
S012/
  ...

Each .npz file corresponds to one subject/trial/step window. Every trial also has three muscle files (see Muscle Activations and Forces).

Arrays

Each file contains:

Key	Shape	Description
subject	scalar string	Subject ID, for example S011.
trial	scalar string	Trial name, for example Trial01.
step_label	scalar string	Step label, currently Step01_Step02.
timestep	(T,)	Zero-based integer timestep over the kinematics sequence.
time_sec	(T,)	Kinematics time in seconds, shifted so the first frame is 0.0.
source_time_sec	(T,)	Same zero-start kinematics time as time_sec.
original_kinematics_time_sec	(T,)	Original OpenSim kinematics timestamps before zero-shifting.
id_source_time_sec	(N_id,)	Original ID timestamps shifted relative to the first kinematics frame.
original_id_source_time_sec	(N_id,)	Original OpenSim ID timestamps before zero-shifting.
joint_angle	(T, N_angle)	Joint angle / coordinate values from the Gait120 kinematics, on the kinematics timeline. Degrees for rotations, metres for pelvis translations.
joint_angle_columns	(N_angle,)	Column names for joint_angle.
joint_torque	(T, N_torque)	ID torques aligned onto the kinematics timeline and divided by subject_weight_kg, in the same sign convention as joint_angle. Values are NaN wherever that column is not valid (see Processing Notes).
joint_torque_columns	(N_torque,)	Column names for joint_torque.
joint_torque_valid	(T,)	Boolean mask of rows where every joint_torque column is valid.
joint_torque_column_valid	(T, N_torque)	Boolean per-column validity. Use it to keep a leg's torques when only the other foot is unmeasured.
joint_torque_sign_convention	scalar string	Sign convention of joint_torque.
joint_torque_normalization	scalar string	Normalization applied to joint_torque, currently divided_by_subject_weight_kg.
joint_torque_units	scalar string	Normalized torque units, currently original_ID_units_per_kg.
subject_weight_kg	scalar float	Subject weight used for normalizing joint_torque.
subject_body_weight_n	scalar float	Subject body weight in Newtons, computed as subject_weight_kg * 9.80665.
computed_grf	(T, N_grf)	GRF values used by ID, interpolated onto the kinematics timeline. Force-vector columns are divided by subject_body_weight_n; COP and torque columns remain in physical units.
computed_grf_columns	(N_grf,)	Column names for computed_grf.
computed_grf_valid	(T,)	Boolean mask where computed_grf has valid values.
computed_grf_normalization	scalar string	Normalization applied to computed_grf, currently force_columns_divided_by_subject_body_weight_n.
computed_grf_units	scalar string	Force columns are in body weight units; COP and torque columns remain in physical units.
qc_notes	scalar string	Quality-control flags for this trial; empty when there are none.
source_files	(3,)	Source kinematics, ID torque, and GRF file paths used for export.

Muscle Activations and Forces

File	Contents
*_muscle_activation.sto	Activations (0–1) of the 80 Rajagopal2016 muscles.
*_muscle_force.sto	Muscle forces (N), then reserve and residual actuators, then the applied external loads.
*_muscle_activations.png	Eight key muscles, right vs left.

• Method: OpenSim static optimization with the same scaled model, joint angles and GRFs as the inverse dynamics. The cost is the sum of squared activations, with force-length-velocity on.
• Actuators added:
  • strong pelvis residuals;
  • torque actuators for the lumbar and arms, which have no muscles in this model;
  • weak (1 Nm) leg reserves, so muscles carry the leg moments.
• Time window and NaN frames: the whole trial, on the .npz timeline. A leg's muscles and reserves are NaN where that leg's torques are invalid. The PNGs shade those frames grey.
• Muscle names are Rajagopal2016 names, e.g. glmed1_r, recfem_r, gasmed_r. Earlier versions of this dataset used gait2392 names, e.g. glut_med1_r, rect_fem_r, med_gas_r.
• Rigid-tendon effect: static optimization assumes a rigid tendon, which tends to overstate plantarflexor activation at push-off. Gastrocnemius often reaches 1.0 there.
• Muscle strength is the generic Rajagopal2016 value and is not scaled to body mass.
  • At about 78–84 % of stance, just before toe-off, the hip flexors and abductors saturate in 44 of 101 subjects, mostly the heavier ones.
  • There, the leg reserves (reserve_* columns) supply part of the hip moment. Their median peak is under 1 Nm, but the worst 5 % of trials reach about 40 Nm.
  • Check reserve_* before relying on hip muscle forces in late stance.
• EMG check. Compared with the measured Gait120 right-leg EMG over 838 gait cycles:
  • r = 0.68–0.90 for tibialis anterior, gastrocnemii, vasti and hamstrings;
  • r ≈ 0.5 for soleus and peroneus longus, which peak ~10 % of the cycle after EMG;
  • rectus femoris does not match (r = −0.22).

Policy Comparison

Run policy inference using joint_angle. The policy output should have the same first dimension T.

Compare predicted normalized torque against OpenSim ID only where joint_torque_valid is true:

import numpy as np

data = np.load("S011/S011_Trial01_Step01_Step02.npz")

joint_angle = data["joint_angle"]
target_torque = data["joint_torque"]
valid = data["joint_torque_valid"]
subject_weight_kg = float(data["subject_weight_kg"])

predicted_torque = policy(joint_angle)
predicted_torque = predicted_torque / subject_weight_kg

error = predicted_torque[valid] - target_torque[valid]

Rows where joint_torque_valid is false have at least one invalid column. To compare one leg's torques while the other foot is down off the force plates, mask per column with joint_torque_column_valid:

col_valid = data["joint_torque_column_valid"]
error = np.where(col_valid, predicted_torque - target_torque, np.nan)