Blame for DETAILS.md

# ForceBody: dataset details

Release 1.3.0 (2026-10-29). The documentation was updated on 2026-11-01 and on 2026-11-02 (CHANGELOG.md). No trial, model or table file changed, except the description column of SCHEMA_v1.3.csv (2026-11-02).

ForceBody pairs the SKEL body model with measured ground reaction forces (GRF), inverse-dynamics joint torques and Monte Carlo torque uncertainty for 10,386 motion trials (9,689,808 frames, 26.89 h, duration = `num_frames` / `fps`) of 140 persons (75 male, 55 female, 10 sex not reported) from 10 studies (AddBiomechanics). Every trial is one NumPy `.npz` file. You do not need nimblephysics, OpenSim or AddBiomechanics tooling to read it.

Since version 1.2, ForceBody ships two SKEL pose sets per trial (one fitted to the kinematics-only AddBiomechanics pass, one to the force-aware dynamics pass), cleaned GRF, recomputed torques and uncertainty for every trial, one body shape per person, corrected metadata, frame-level quality masks, and fixed splits. Read "Choosing a pose source" and "Leakage policy" before training anything that predicts or scores forces.

| study | persons | trials |
|---|---|---|
| Carter2023 | 50 | 7,006 |
| Falisse2017 | 11 | 70 |
| Fregly2012 | 4 | 78 |
| Hamner2013 | 10 | 40 |
| Han2023 | 7 | 240 |
| Li2021 | 1 | 65 |
| Moore2015 | 12 | 1,019 |
| Tiziana2019 | 25 | 619 |
| Uhlrich2023 | 10 | 160 |
| vanderZee2022 | 10 | 1,089 |
| total | 140 | 10,386 |

Study ids follow the first author and year of the source study (`study_citation` gives the citation key, Tiziana2019 = Lencioni et al. 2019). Release 1.3.0 renamed Hammer2013 to Hamner2013 (Hamner and Delp 2013). The 40 trial keys, folders and person ids changed (`uid_map_v12_to_v13.csv`, manifest column `trial_uid_v12`). Fields that point into the AddBiomechanics source (`b3d_path`, `subject_id_raw`, `trial_uid_v1`, the `models/` file names) keep the AddBiomechanics folder spelling Hammer2013.

## Layout

This Hugging Face repository holds the data only. The loader (`forcebody_io.py`), the renderer (`visualize.py`, `VISUALIZATION.md`), the benchmark card, cache builder and scorer (`benchmark/`) and the baseline code are in the GitHub repository https://github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026. The v1.1 provenance files and the Monte Carlo torque percentiles are not part of this upload.

Downloaded as in the Quick start of the dataset card (`--local-dir ForceBody`), the folder holds:

```
ForceBody/
  trials/<study>/S<NN>/<trial_uid>.npz        core arrays, 547 fields (SCHEMA_v1.3.csv)
  manifest.csv, manifest_train.csv, manifest_test.csv   one row per trial (every scalar field except the four JSON records
                                              head_fit_json*, skel_fit_metrics_json*, + metadata)
  uid_map_v12_to_v13.csv                     the 40 trial keys and paths renamed in 1.3.0 (Hammer2013 -> Hamner2013)
  name_map.csv, name_map.json                trial_uid -> AddBiomechanics source file and trial (`split` = the AddBiomechanics source split, not split_addb)
  persons.csv                                one row per person: betas, sex, height, masses, SHAPY measures, splits
  subjects.csv                               per-person metadata table (v1.1 layout, n_with_uncertainty is the v1.1
                                              count, see below)
  gait_events.csv                            GRF heel strikes and toe-offs (walking / running)
  models/                                    AddB pass-0 and pass-2 scaled OpenSim models per b3d file (.osim.gz),
                                              models_index.csv (model file -> source b3d file and processing pass),
                                              body_segments.csv (segment masses, local COM)
  marker_vertex_map.json                     marker name -> SKEL skin vertex used by the fit
  SCHEMA_v1.3.csv                            field, dtype, shape, unit, frame, version_added, description
  README.md                                  dataset card
  DETAILS.md                                 this file
  CHANGELOG.md                               version history
  MD5SUMS.txt                                md5 of the data files (with SCHEMA_v1.3.csv) and of README.md,
                                              DETAILS.md, CHANGELOG.md (not croissant.json, LICENSE.md, assets/)
  croissant.json                             Croissant metadata with Responsible AI fields
  LICENSE.md                                 license and attribution terms
  assets/                                    images of the dataset card
```

subjects.csv keeps the v1.1 column layout. Its `n_with_uncertainty` column is the v1.1 count (8,652 trials in total). For release 1.3.0 use `has_uncertainty` in manifest.csv (10,382 trials). Ten subjects.csv columns have no row in SCHEMA_v1.3.csv. `sex_v12` equals `sex`. `n_trials` is the number of trials of the person, `n_b3d_files` the number of AddBiomechanics source files and `n_mass_values` the number of distinct `mass_kg` values. `activities` and `activity_names` list the `activity_name` values, and `activity_classes` lists the `activity_recovered` values ('|'-joined). `usable_frac` and `eligible_eval` do not match the current trial flags. Do not use them to select persons. Use the trial flags and split columns of manifest.csv.

The `.osim.gz` files in `models/` store the internal gzip name `<name>.osim.gz.writing_tmp`. Decompress them with `gunzip` or Python `gzip`. Tools that use the stored name (7-Zip, `gunzip -N`) write `<name>.osim.gz.writing_tmp`. Rename such a file to `<name>.osim`.

Some SCHEMA_v1.3.csv descriptions carry short build tags in parentheses (for example S19 F3 or D20). They are internal issue labels of the build and need no lookup.

## Choosing a pose source

Each trial has two SKEL fits. Both have every field and every QC mask. Pick one with `pose_source`.

| `pose_source` | fields | fitted to | default use |
|---|---|---|---|
| `pass0` (alias `kin`) | `poses`, `trans`, `joints`, `q_kin`, `qd_kin`, `qdd_kin`, `addb_joints_kin` | AddB kinematics pass (pass 0) + the 30 Hz 2nd-order Butterworth filter that AddB applies for pass 1 (see below). Never saw the force plates | force and torque estimation, physics critics p(tau \| motion), any metric that scores forces from motion |
| `pass2` (alias `dyn`) | `poses_dyn`, `trans_dyn`, `joints_dyn`, `q_gt`, `qd_gt`, `qdd_gt`, `addb_joints_dyn` | AddB dynamics pass (pass 2). Consistent with the GRF and torque labels | text-to-motion (motion only, or motion + GRF + torque), T2M evaluators |

```python
# run in ForceBody-NeurIPS-ED-2026/code (Quick start step 3 of the dataset card)
import forcebody_io as fio
d = fio.load_trial("../../ForceBody/trials/Moore2015/S10/Moore2015_S10_walk_014.npz", pose_source="pass0")
# common keys: poses, trans, joints, q, qd, qdd, addb_joints, pose_usable, pose_source
# plus every GRF / torque / mask / metadata field of the file
```

`pose_source` is required (`ValueError` otherwise). `pose_usable` is always True for `pass0` and equals `poses_dyn_usable` for `pass2` (False in 15 trials: force/motion-capture desync). What pass 2 changes, from the AddB documentation and our audit. AddB moves the pelvis so the centre-of-mass path matches the measured GRF, then refits poses and body scales. COM-GRF Newton consistency improves (r 0.9947 -> 0.9998), but stance-foot sliding rises (0.064 -> 0.079 m/s) and the p90 marker error rises (26.8 -> 36.0 mm). Joint centres differ by a median 9.6 mm, mostly a constant bone-rescaling offset. Pass 2 also changes the motion on frames without valid GRF (`grf_valid_mask` False on 1,348,203 of the 9,689,808 release frames, 13.9%), except in trials with `qc_pass2_equals_pass1`.

`q_kin` is AddB pass 0 filtered with the same 2nd-order 30 Hz zero-lag Butterworth filter that AddB applies for pass 1. `q_kin` reproduces the stored AddB pass 1 in 8,218 trials (`q_kin_pass1_match`). `q_kin_pass1_mismatch_reason` gives the reason for the other 2,168 trials (for example the arm re-solve of 875 trials, see the next section).

Re-timed trials (27, `skel_fit_target_dyn == "kin_retimed"`). The force stream was shifted in time to match the motion capture (`grf_retimed_ms`). Their pass-2 fit had been bent to the misaligned force, so `poses_dyn`, `trans_dyn`, `joints_dyn` and `addb_joints_dyn` equal the pass-0 fields, `q_gt`, `qd_gt`, `qdd_gt` and `tau_gt_det` are not usable (`q_gt_usable`, `tau_gt_det_usable` False), and `tau_selfconsistent`, `mu_tau` and `sigma_tau` are computed on `q_kin` with the pass-2 skeleton (`tau_sc_kinematics`, `mc_kinematics`). `load_trial(..., "pass2")` returns `q_kin` as `q` for these trials (`q_source`).

## SKEL fit versions and fix rounds

Both pose sets come from one refit code base that went through several fix rounds before this release. `skel_fit_code_version` / `skel_fit_code_version_dyn` and `skel_fit_source_run` / `skel_fit_source_run_dyn` record per trial which code produced each fit.

| fit | code version | trials |
|---|---|---|
| `poses` | v1.2-fixround-2026-10-03c | 10,291 |
| `poses` | v1.2-fixround-2026-10-04d | 95 |
| `poses_dyn` | v1.2-fixround-2026-10-03c | 8,940 |
| `poses_dyn` | v1.2-fixround-2026-10-07e | 9 |
| `poses_dyn` | v1.2-fixround-2026-10-08i | 815 |
| `poses_dyn` | v1.2-fixround-2026-10-09l | 622 |

What the rounds changed (all trials unless a count is given):

- scapula: the right scapula elevation bound of SKEL `kin_skel.py` has the wrong sign. It is replaced by the mirror of the left bound ([0.1, 0.4] rad). The bound sizes (0.1, 0.4 rad) have no published source (SKEL issue #19) and elevation sits on them in many fits (see Known issues).
- marker handling: aliases for trunk / shoulder markers (Han2023, Uhlrich2023) and acromion vertices moved to the acromion tip (`marker_vertex_map.json`).
- kinematics target (`q_kin`): where the AddB pass-0 left arm parks on its joint limit or hops (> 30 deg per frame) and where the elbow parks at 0, the arm is re-solved by inverse kinematics to the markers before the 30 Hz filter. A window is kept only if the marker error does not rise (hop windows: by at most 1 mm). 875 trials (`q_kin_source == "addb_pass0_armfix_lp30"`). Inside full-body hop windows the leg and trunk target also changes.
- fit: a continuity guard repairs fit-only branch flips of arm, leg and root DOF groups. Stage-2 joint weights of femur and tibia raised from 5 / 3 to 10 (our choice) so the knee follows the target. An extra lower-body stage (2f) is kept when it passes the whole-body error gate. Trade-off: the pelvis joint error rises (median about 2 mm) because the SKEL femur is shorter than the AddB femur.
- head: the head is pinned during the main fit and fitted afterwards to the head markers (orientation only, neck limits of bsm.osim). 1,948 trials have `head_postfit`. The others (`qc_head_untracked`) keep head DOFs 23-25 at 0, so the head follows the thorax.

### Dynamics-pass target repair (`poses_dyn` is not pure pass 2 everywhere)

`poses_dyn` is fitted to the AddB dynamics pass (`q_gt`) except where that target was repaired. `skel_fit_target_dyn` names the target, `dyn_target_source [T,37]` marks per frame and AddB DOF which values replaced `q_gt` (0 = pure pass 2), `dyn_target_modified_frac` is the fraction of modified frames, and `addb_joints_dyn` are the joint centres of the target actually fitted. `q_gt` itself is unchanged, so in the flagged frames `poses_dyn` and `q_gt` disagree.

| `skel_fit_target_dyn` | trials | target |
|---|---|---|
| `addb_pass2` | 8,939 | pure pass 2 (`q_gt`) |
| `addb_pass2_armsplice` | 9 | pass 2 with the kinematics-target (re-solved) arm spliced in inside the accepted arm windows, cosine ramps |
| `addb_pass2_repaired_v3a` | 791 | pass 2 with the kinematics target (whole body, aligned at the window edges) in frames without valid GRF, the kinematics root on GRF-valid frames where the pass-2 pelvis deviates by more than 50 mm, kinematics values on pass-0 IK outlier frames, left-arm splice. 0.25 s cosine ramps |
| `addb_pass2_repaired_v4s` | 620 | v3a plus a position-based left-arm IK on the pass-2 skeleton and the AddB acceleration-minimizing filter on all 37 DOFs (removes 6-30 Hz wobble of pass 2) |
| `kin_retimed` | 27 | `poses_dyn` = `poses` (re-timed trials, below) |

Bits of `dyn_target_source`: 1 kinematics value (GRF-invalid window or pass-2 root deviation), 2 left-arm splice, 4 pass-0 IK outlier repair, 8 left-arm position IK, 16 acceleration-minimizing filter, 32 other difference from `q_gt` (arm splice). 1,392 trials have a modified frame (median modified fraction 0.83). 1,048,169 of 9,689,808 frames. `dyn_arm_spliced [T,2]` marks frames whose dyn-target arm comes from the kinematics or IK arm (622 trials, `qc_dyn_arm_spliced`). The repair trades Newton consistency for a plausible root and arms where pass 2 had no force support: on the repaired trials the root path follows the kinematics target in GRF-invalid windows, so GRF-free frames of `poses_dyn` are no longer force-aware there. For a force-aware motion source with the arm and root fixes, use `poses_dyn` with its masks. For pure AddB pass-2 kinematics use `q_gt`.

## Leakage policy

AddBiomechanics warns, "If you use the output of the dynamics pass, you will be training your model on data that has already been fine tuned to match the force plate data, and your model will overfit to the clues that the optimizer inserts into the motion", and for GRF prediction to "use as input features only the kinematics pass" (nimblephysics documentation, working with AddBiomechanics data). GaitDynamics uses the kinematics pass "to prevent the leakage of force data ... into the kinematics". InferBiomechanics (AddB benchmark code) takes inputs from the kinematics pass and labels from the last pass.

The effect is large here. Recovering the total GRF from the centre-of-mass acceleration alone, with no learning, has a median error of 2.56 N/kg on pass-1 kinematics and 0.2 N/kg on pass-2 kinematics.

The rule is an allowlist. A model that predicts or scores forces or torques may take as input only the pass-0 pose set (`poses`, `trans`, `joints`, `q_kin`, `qd_kin`, `qdd_kin`, `addb_joints_kin`, or `load_trial(..., "pass0")`), `betas`, `skel_gender`, `mass_kg`, `height_m`, `sex` and `age` (these fields are `forcebody_io.ALLOWED_INPUT_FIELDS`), and person-free text metadata. Every other field is excluded, in particular the dynamics-pass pose set (`poses_dyn`, `trans_dyn`, `joints_dyn`, `addb_joints_dyn`, `q_gt`, `qd_gt`, `qdd_gt`), the labels (`tau_*`, `mu_tau`, `sigma_tau`, `percentiles`, `grf*`, `cop*`, `contact_*`, `grf_valid_mask`) and the fields derived from the GRF: `mass_kg_grf`, `mass_kg_grf_stand`, `mass_dyn_scale`, `mass_dyn_used_kg`, `mass_kg_dynamics*`, `cadence_spm_grf`, `stride_time_s_med`, `duty_factor_*`, `gait_events.csv`, `support_side`, `support_side_grf_share`, `caption_support_side`, `floor_height_m`, `floor_tilt_*`, `speed_belt_grf_mps`, `addb_belt_speed_mps`, `speed_mps` (running), `tilt_deg` and the frame array `chair_contact`, and the masks and flags `mc_frame_ok`, `kin_frame_ok`, `eval_*`, `t2m_*`, `s8b_*`, the usability flags (for example `q_gt_usable`, `usable_provisional`), the `qc_*` fields and every `*_dyn` / `dyn_*` field (`dyn_target_source` is placed by the GRF-valid windows). `forcebody_io.LEAKY_INPUT_FIELDS` is only a safety net. It flags many of these fields, but it is not guaranteed to be complete. The allowlist above defines the valid inputs. The benchmark cache builder refuses every input that is not a benchmark input. Results computed with `pass2` inputs must be labelled leaky and are not comparable with the official benchmark numbers. The body shape (`betas`) was fitted to pass-0 bone lengths, the recorded height and a person mass target. For Moore2015 that target is the GRF calibration-pose mass (a static, person-level weight).

## Force benchmark

The force benchmark (joint torques T-tau and per-foot GRF T-GRF from motion) is defined in `code/benchmark/BENCHMARK.md` of the GitHub repository https://github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026 with the machine-readable `benchmark/spec_v12.json`, a cache builder (`benchmark/build_bench_cache_v12.py`) and the reference scorer (`benchmark/score_v12.py`). The scripts import `forcebody_io.py` from the `code/` folder of the GitHub repository. The baselines retrained on 1.2.1 (three seeds, the numbers of the camera-ready paper, see Paper numbers) are in the Benchmark section of the GitHub README. The benchmark data of 1.3.0 are identical to 1.2.1 except the 40 renamed Hamner2013 keys.

## Conventions

| item | convention |
|---|---|
| frames | `addb_lab`: raw AddB lab frame, y up (grf, cop, contact_torque, markers, q_*, tau_*). `skel_world`: `flip_xz(addb_lab)` (x -> -x, z -> -z, y unchanged): poses, trans, joints, addb_joints_*, *_aligned, cop_aligned_foot. Every `*_aligned` field is exactly `flip_xz` of its raw field (max difference 0) |
| gravity | -y in every study |
| contacts | index 0 = right foot (calcn_r), 1 = left foot (calcn_l), 2 = pelvis (Falisse2017 chair) |
| trans | SKEL root translation (`trans_frame == "skel_world"`). The pelvis joint is `trans + j0(betas)`, not `trans`. Here j0 is about (0.0095, -0.236, 0.103) m for a male rest pose, so rotating `trans` alone misplaces the body by 0.12-0.16 m |
| joints | `joints[T,24,3]` in `joint_names` order (SKEL / OpenSim: 0 pelvis, 1 femur_r, ...) |
| DOF names | `skel_pose_names` (46 SKEL pose DOFs. `skel_dof_valid` False = pinned to 0 in the main fit, namely head DOFs 23-25 and wrist DOFs 34, 35, 44, 45. The head post-fit then sets the head DOFs where `head_postfit` is True (1,948 trials). The wrist DOFs stay 0), `dof_names` = `dof_names_opensim` (37 AddB DOFs, `<OpenSim joint>_<coordinate index>` names, see below) |
| time | `t_i = i / fps_true_grid`. `fps` is the nominal rate. The true grid differs slightly (61 Hz: 64, 99 Hz: 470, 100 Hz: 9,640, 101 Hz: 76, 201 Hz: 111, 250 Hz: 25 trials by rounded rate) |
| heading | use `heading_canonical_deg` (SKEL frame) only |
| mirroring | swapping left/right SKEL DOFs with a sign table is not an exact mirror (MPJPE 32 mm, feet 45-77 mm). Mirror the mesh or the joint positions instead |

AddB DOF order. `dof_names` and `dof_names_opensim` hold the same 37 names. They label the columns of `q_*`, `tau_*`, `mu_tau` and `sigma_tau`. The names are `<OpenSim joint>_<coordinate index>` (for example `hip_r_0`, `walker_knee_r_0`), not OpenSim coordinate names. Index k is the k-th coordinate of that joint in the AddB OpenSim model (`models/`, the Rajagopal coordinate set):

| columns | names | OpenSim coordinates |
|---|---|---|
| 0-5 | `ground_pelvis_0`-`5` | pelvis_tilt, pelvis_list, pelvis_rotation, pelvis_tx, pelvis_ty, pelvis_tz |
| 6-12 | `hip_r_0`-`2`, `walker_knee_r_0`, `ankle_r_0`, `subtalar_r_0`, `mtp_r_0` | hip_flexion_r, hip_adduction_r, hip_rotation_r, knee_angle_r, ankle_angle_r, subtalar_angle_r, mtp_angle_r |
| 13-19 | same for the left leg (`_l`) | same coordinates, `_l` |
| 20-22 | `back_0`-`2` | lumbar_extension, lumbar_bending, lumbar_rotation |
| 23-29 | `acromial_r_0`-`2`, `elbow_r_0`, `radioulnar_r_0`, `radius_hand_r_0`-`1` | arm_flex_r, arm_add_r, arm_rot_r, elbow_flex_r, pro_sup_r, wrist_flex_r, wrist_dev_r |
| 30-36 | same for the left arm (`_l`) | same coordinates, `_l` |

In `tau_selfconsistent`, `tau_gt_det`, `mu_tau` and `sigma_tau`, columns 0-2 are root residual moments (N m) and columns 3-5 are root residual forces (N). Columns 6-36 are joint torques (N m).

## Body mass fields

| field | meaning | use |
|---|---|---|
| `mass_kg` | recorded mass (unchanged from v1.1) | model and critic inputs |
| `mass_kg_dynamics` | AddB pass-2 skeleton mass times `mass_dyn_scale` (already applied) | the only mass for normalizing GRF, torque and sigma labels (N/kg, N m/kg, BW) |
| `mass_kg_dynamics_v11`, `mass_dyn_scale` | v1.1 dynamics mass and the approved per-file rescale | provenance. Never multiply `mass_kg_dynamics` by `mass_dyn_scale` again. `mass_kg_dynamics_v11` stays in the core file (the only `_v11` value there besides the QC count `qc_grf_endpoint_valid_v11`) |
| `mass_kg_grf` | per-trial GRF-implied mass estimate | diagnostic |
| `caption_mass_kg`, `shape_mass_target_kg`, `mass_source` | person mass for captions and for the shape fit (recorded. Moore2015 = GRF calibration pose) | text, shape |

`mass_dyn_scale` differs from 1 in 491 trials (14 b3d files), where the AddB link masses were inconsistent with the measured GRF.

## Quality masks

Nothing is deleted. Problems are flagged. Frame masks (`[T]` or `[T,2]`): `grf_frame_ok` (GRF validity: GRF losses and statistics), `grf_frame_strict_ok` (whole frame without zero-force or misassigned feet), `grf_foot_ok` (per foot, columns right, left, 1.2.1), `tau_frame_ok` (torque losses, use `eval_tau_frame_ok` for evaluation), `mc_frame_ok` (use of sigma_tau, equal to `tau_frame_ok` and `kin_frame_ok` and `has_uncertainty`), `kin_frame_ok` (pelvis not flipped), `cop_frame_ok` (per foot), `grf_contact_state` (per foot: 1 = in contact on a `grf_frame_ok` frame, 0 = not in contact on a `grf_frame_ok` frame, -1 = unknown: no measured force, or a measured frame outside `grf_frame_ok`), `eval_tau_frame_ok` (= `tau_frame_ok` and `kin_frame_ok`), `eval_grf_frame_ok` (= `grf_frame_ok` and `kin_frame_ok`).

GRF validity and CoP trust (1.2.1). `grf_frame_ok` is GRF validity only: `grf_valid_mask`, not an endpoint or a blanked tail, and a consistent force label on every used contact (`grf_contact_ok`). A loaded contact with `contact_status` 3 (force valid, CoP projected on the foot and not trusted) counts as valid. CoP trust gates only `cop_frame_ok` (trusted CoP status 1, 2 or 9 on that foot) and `tau_frame_ok` (no used contact with status 3, because the torque depends on the CoP). In 1.2.0 the status-3 frames were not in `grf_frame_ok`, so on fixed-point-CoP trials (`qc_cop_fixed_point`, Carter2023) the GRF masks kept only unloaded frames (CHANGELOG, erratum). In 1.2.1 170,158 status-3 frames of 1,158 trials are in `grf_frame_ok`. The QC scalars are still computed on `grf_frame_ok` without the status-3 frames, as in 1.2.0, so their values did not change, except the mask summaries (`grf_frame_ok_frac`, `grf_frame_strict_ok_frac`, `qc_grf_contact_bad_frames`) and the evaluation flags.

Per-foot GRF mask (1.2.1). `grf_frame_strict_ok` drops the whole frame when either foot is zero-force grounded, force-in-air, unmeasured in stance or has its CoP off the foot. `grf_foot_ok` drops only the zero-force-grounded foot itself. The other three flags point to a likely plate misassignment (the force of the other foot may be on the wrong plate), so they still drop both feet of the frame (`grf_foot_ok.all(1)` equals `grf_frame_strict_ok`). It keeps 854,888 loaded foot-frames (vertical GRF > 20 N) that the strict mask removes.

Masked values are not NaN. Outside a mask the arrays keep their raw values. `cop` and `contact_torque` are 0 where the force is 0, the GRF of the AddB endpoint frames is 0, and `grf` / `grf_aligned` can be non-zero on frames outside `grf_valid_mask` (force readings that AddB marked as missing). A 0 is not a measurement, so apply the mask.

Trial flags:

| flag | trials True |
|---|---|
| `force_label_usable` | 10,372 |
| `torque_label_usable` | 10,067 |
| `eval_motion_trial_ok` | 9,697 |
| `eval_grf_trial_ok` | 9,769 |
| `eval_torque_trial_ok` | 8,256 |
| `qc_pass2_equals_pass1` | 1,249 |
| `qc_skel_grf_newton_mismatch` | 244 |
| `qc_grf_single_channel` | 37 |
| `qc_mass_dyn_vs_grf_flag` | 20 |
| `qc_overlap_pair_drop` | 8 |
| `qc_overlap_motion_drop` | 200 |

Evaluation flags. `eval_grf_trial_ok` and `eval_torque_trial_ok` require usable force labels and exclude single-channel force plates (`qc_grf_single_channel`: one plate carries both feet, so per-foot GRF and the foot assignment of the torques are wrong, also excluded from `torque_label_usable`), byte-identical duplicates (`qc_duplicate_drop`), the Han2023 trimmed copies of a re-exported capture (`qc_overlap_pair_drop` = `eval_exclude_duplicate` and not `qc_duplicate_drop`), static calibration, transition and pause segments, and trials whose dynamics mass disagrees with the GRF-implied mass by more than 5% (`qc_mass_dyn_vs_grf_flag`). `eval_torque_trial_ok` also requires a root residual force of at most 0.1 BW. Since 1.2.1 both also require at least one loaded evaluation frame (`qc_grf_loaded_eval_frames` > 0: a foot vertical GRF above 20 N on an `eval_grf_frame_ok` frame, for torques also `qc_tau_loaded_eval_frames` > 0). 83 trials with usable force labels have none (most have `grf_valid_mask` False on every frame). `eval_motion_trial_ok` applies the same duplicate / segment exclusions and in addition drops one member of each overlapping Tiziana2019 capture pair (`qc_overlap_motion_drop`: the member with fewer GRF-valid frames). The Tiziana2019 pairs share no GRF-valid frame, so both members stay in force evaluation. Always combine a trial flag with its frame mask.

`force_label_usable` is False where force and motion capture are out of sync and could not be re-timed (14 trials). Every force, torque, MC, contact-state and CoP frame mask is then False. `grf_contact_ok`, the per-contact label-consistency input of `grf_frame_ok`, is not cleared. Never use it without `grf_frame_ok`. `qc_skel_grf_newton_mismatch` marks trials whose SKEL root path does not follow the GRF without any time-lag evidence. Their force labels are kept. Pathological persons are never excluded. `eval_stratum` reports them separately. Every QC field is described in SCHEMA_v1.3.csv.

Falisse2017 chair channel (contact index 2, `contact_bodies` = `pelvis`). `chair_contact` [T] marks the frames where the subject sits on the instrumented chair (the `grf_contact_label` hysteresis on the chair vertical force, not restricted to `grf_frame_ok`, most seated frames are outside `grf_valid_mask`), `qc_chair_channel_active` marks trials with any such frame, and `qc_extra_channel_zero` marks trials whose chair channel carries no force at all (walk, stairs). On those trials `grf_contact_ok[:, 2]` is False on every frame and `grf_frame_ok` ignores the channel. The trial-level manifest column `chair_contact` (protocol metadata) is a different field from the per-frame npz array of the same name.

### Pose masks of the fix rounds

These masks describe the SKEL pose sets (True = usable). Each exists for both pose sets (`_kin` = `poses`, `_dyn` = `poses_dyn`). AND them into the loss mask of the listed DOFs. Masks are dilated by 0.1 s (our choice).

| field | shape | False on | trials with a False frame | frames |
|---|---|---|---|---|
| `arm_event_ok_kin` | [T,2] (left, right) | arm discontinuities (arm-jump diagnosis of 2026-10-07 and a re-scan of the final fits). Arm DOFs | 272 | 231,278 |
| `arm_event_ok_dyn` | [T,2] | same for `poses_dyn` | 316 | 217,726 |
| `trunk_event_ok_kin` | [T] | AddB pass-0 trunk shake (class A3). Lumbar / thorax DOFs 17-22 | 143 | 200,423 |
| `trunk_event_ok_dyn` | [T] | same for `poses_dyn` | 155 | 200,116 |
| `dyn_root_jump_ok` | [T] | pass-2-only whole-body jumps (class A6): drop the whole `poses_dyn` frame | 108 | 2,858 |
| `fit_event_ok_kin` | [T,7] | temporal events of the final fit per joint group (`fit_event_groups`: pelvis, leg_r, leg_l, spine, head, arm_r, arm_l): fit-only spikes, spikes inherited from the target, one-frame DOF steps, and Euler branch flips (one-frame step > pi) | 1,820 | 251,010 |
| `fit_event_ok_dyn` | [T,7] | same for `poses_dyn`, measured against its own (repaired) target | 1,808 | 254,087 |
| `kin_hop_edge_frame_ok_r` | [T] | right-arm frames at a hop-window edge of the arm re-solve where two Euler branches are joined (`qc_kin_hop_edge_branch_blend`) | 3 | 75 |

Also per trial: `addb_limit_mask_kin` / `addb_limit_mask_dyn` [T,11] (the fit target sits on an AddB joint limit, the dyn mask is the union of the pass-2 and pass-0 on-limit frames, `addb_limit_mask_names`), `qc_arm_event_classes_kin` / `_dyn` (diagnosis classes present), and the trial flags below. Where `qc_upper_body_unobserved` is True (no torso / arm marker at all, e.g. vanderZee2022 S09) the arm, trunk and head masks are False on every frame of both pose sets: the SKEL arms, trunk and head are unconstrained there.

| trial flag | trials True | meaning |
|---|---|---|
| `qc_upper_body_unobserved` | 133 | no upper-body marker. Arms / trunk / head masked |
| `qc_head_untracked` | 8,438 | fewer than 2 head markers. Head DOFs 23-25 are 0 |
| `qc_kin_armfix_edge` | 20 | arm re-solve window-edge artefact of the kinematics target (frames in `arm_event_ok_kin`) |
| `qc_kin_hop_edge_branch_blend` | 3 | hop-window edge joins two right-arm Euler branches |
| `qc_upper6_rise10_kin` | 13 | upper-body joint error of `poses` rose by > 10 mm in the fix rounds (our threshold). Arm-marker error is better in almost all. Flag, not exclusion |
| `qc_upper6_rise10_dyn` | 18 | same for `poses_dyn` |
| `qc_heading_rms3_kin` | 3 | pelvis heading lost > 0.1 correlation and detrended heading RMS > 3 deg (our thresholds) |
| `qc_heading_rms3_dyn` | 2 | same for `poses_dyn` |
| `qc_dyn_arm_spliced` | 622 | `poses_dyn` arms follow the kinematics / IK arm in `dyn_arm_spliced` frames |
| `qc_tau_sc_q_kin_pre_armfix` | 0 | re-timed trial whose torque labels were computed on `q_kin` before the arm re-solve (`mc_kinematics == "q_kin_pre_armfix"`) |
| `qc_pose_wrap_n_kin` > 0 | 1 | Euler branch flip in `poses` (masked, not unwrapped) |
| `qc_pose_wrap_n_dyn` > 0 | 1 | same in `poses_dyn` |
| `qc_head_at_limit_frac_*` > 0.05 | 11 | head post-fit at a neck limit on > 5% of frames. Do not use the head orientation there |

Added in 1.2.1 (T2M audit, flags and masks only, no pose refit):

| field | trials | meaning |
|---|---|---|
| `qc_q_kin_wrap` | 68 | `q_kin` (fit target of `poses`) beyond 2 pi on a shoulder DOF (an AddB Euler branch. `qc_q_kin_abs_gt_2pi_dofs`). The frames of that arm are False in `arm_event_ok_kin`, and in `arm_event_ok_dyn` on kin_retimed trials. Not unwrapped |
| `knee_clamp_frame_kin` / `_dyn` [T,2] (right, left), `qc_knee_clamp_frac_*` | 4,870 (kin, any frame). 1,061 kin / 1,294 dyn above 10% of frames | SKEL knee on its 0 bound while the AddB target knee is flexed 3 deg or more (our threshold). Flag only. `forcebody_io.load_trial` returns the mask of the chosen set as `knee_clamp_frame` |
| `qc_belt_speed_change` / `qc_belt_speed_change_suspected` | 4 / 8 | treadmill belt labelled constant whose speed changes in the trial (both pose sets / one pose set) |
| `label_conflict_codes` contains SEGMENT_MOTION_MISMATCH | 28 | segment label and motion disagree (Carter2023 static-trial stand segments with a leg lift, Han2023 pause / sit / short action segments). Labels and captions not changed |

## Uncertainty

`mu_tau` and `sigma_tau` ship in the core file for all 10,386 trials, the torque `percentiles` (with `percentile_values`) are kept in the internal v1.1 provenance archive and are not part of this upload. `has_uncertainty` is True for 10,382 (False: the trial is shorter than one period of the 6 Hz noise band, sigma is 0). `has_uncertainty_v11` (manifest) marks the 8,652 trials that had MC in the version of the submitted paper. sigma_tau is the kinematic sensitivity of the torque to 1.6 deg band-limited joint-angle noise on `q_gt` (pass 2, `q_kin` on re-timed trials) with the measured GRF held fixed. It is not the uncertainty of `poses`, and not a label-reliability score. Normalize it by `mass_kg_dynamics` only, and use it only where `mc_frame_ok` and `torque_label_usable` hold. Per-trial ratios to the sigma of the submitted-paper version have p5-p95 0.85-1.20. The sigma statistics of the paper use the trials with `has_uncertainty` and `grf_valid_frac` >= 0.5 (8,732 trials).

## Time alignment

| item | status |
|---|---|
| release grid | one grid per trial for kinematics, forces and MC (`fps_true_grid`) |
| AddB endpoints | frames 0 and T-1 carry no GRF. They are never in `grf_frame_ok` |
| duration | `duration_s` = `num_frames` / `fps` (nominal rate, v1 convention). The time span of the grid is (`num_frames` - 1) / `fps_true_grid`, about one frame shorter |
| force/mocap re-timing | 27 trials re-timed and validated against independent marker-based gait events. Applied shift median 277 ms (range -85 to 320 ms) |
| partial invalidation | 1 trial: force invalid from `force_invalid_from_frame` |
| lag diagnostics | `qc_grf_kin_toeoff_lag_ms`, `qc_grf_kin_onset_lag_ms`, `qc_grf_newton_seg_lag_ms`, `qc_cop_foot_shift_ms` (positive = kinematics later than force) |

## Floor and gait events

- `floor_height_m`: median centre-of-pressure height (SKEL frame, y) over trusted firm-stance frames (`contact_status` 1, foot vertical force > 0.5 BW of `mass_kg_dynamics`, `grf_valid_mask`). NaN with fewer than 30 such frames. It comes from the force plates. On stairs and step activities it is the height of the loaded step surface, not of the floor.
- No floor normal is shipped. A plane fitted to the SKEL sole height over the CoP positions tilts by 2-3 deg on level treadmills (heel-to-toe roll). The QC fields `floor_tilt_x_deg` / `floor_tilt_z_deg` (plane through the firm-stance CoPs) and `floor_tilt_x_kin_deg` / `floor_tilt_z_kin_deg` (SKEL sole height at the same samples) are diagnostics only. Protocol inclines are in `incline_grade_pct` / `incline_label`.
- `gait_events.csv`: heel strikes (`grf_contact_state` 0 -> 1) and toe-offs (1 -> 0) of each foot on walking and running trials with usable force labels. `frame` on the release grid, `time_s = frame / fps_true_grid`. `stride_time_s_med` is the median same-foot heel-strike interval (0.4-2.5 s, strides with an unknown (-1) frame inside are not used), `cadence_spm_grf` = 120 / `stride_time_s_med`, `duty_factor_r` / `duty_factor_l` the median stance fraction of a stride.
- Consistency gate: where `cadence_spm_grf` differs from the kinematic `cadence_spm` (metadata) by more than 10% (our threshold), `qc_gait_event_inconsistent` is True, the four GRF gait scalars are NaN and the trial has no rows in gait_events.csv (28 trials).
- Coverage: 8,940 of the 9,700 walking or running trials with usable force labels have rows. The other 760 trials have none. 28 fail the consistency gate. 732 have no 0 -> 1 or 1 -> 0 transition of `grf_contact_state`, among them all 609 Tiziana2019 walking and running trials.
- Carter2023 running. In most trials `grf_valid_mask` also covers the flight frames (median `grf_valid_frac` 1.00). In some trials AddB marks the flight frames as missing GRF, and there GRF labels, T-GRF scores and `grf_contact_state` (-1) cover stance frames only and strides that contain such frames give no GRF gait scalars. On fixed-point-CoP trials (`qc_cop_fixed_point`) every loaded frame has `contact_status` 3. Since 1.2.1 these frames are in `grf_frame_ok` and `grf_contact_state`, so these trials have GRF gait events, but they stay out of `cop_frame_ok` and `tau_frame_ok`. (1.2.0 stated the opposite: its GRF masks kept only the unloaded frames of these trials, CHANGELOG erratum.)

## Metadata

Every metadata column is in manifest.csv and the per-trial ones also in each trial file. SCHEMA_v1.3.csv gives the definition of each column. Groups: identity and capture structure (`person_id`, `subject_base`, `capture_id`, `segment_index`, `n_segments_in_capture`, `segment_role`, `qc_overlap_of`, `eval_exclude_duplicate`), activity (`activity_name`, `activity_recovered`, `activity`, `activity_sequence`, `activity_modifier`, `gait_style`, `source_task`, `setting`, `setting_motion`, `path`, `support_side`), protocol (`protocol_condition`, `condition_key`, `speed_*`, `incline_*`, `belt_*`, `cadence_*`, `step_*`), person (`sex`, `age`, `bmi`, `health_status`, `health_detail`, `footwear`) and captions (`caption_*`). The measured values (`speed_mps`, `cadence_spm`, `step_width_m`) come from forward kinematics of the AddB model. The running treadmill speed uses GRF stance frames, and `support_side_grf_share` and `speed_belt_grf_mps` use the GRF (Leakage policy).

## Body shape

`betas` are fitted once per person (identical in all trials of a person, persons.csv) to the pass-0 bone lengths, the recorded height and the person mass target, and `skel_gender` selects the SKEL model (`skel_gender_source`: height_rule 10, sex_label 130 persons). persons.csv reports SHAPY-style measurements of the shaped mesh (height = |y(HeadTop 411) - y(HeelLeft 3466)|, mass = volume x 985 kg/m^3, chest / waist / hip girths at vertices 6489 / 3501 / 1210). The girths are almost fully determined by height and mass (R^2 0.987-0.994 for a linear fit on the recorded height and the person mass target within each SKEL gender model, and 0.31 for the hip girth when both genders are pooled). 41 (female) / 52 (male)% of the betas variance is bone proportion that height and mass do not explain. Keep claims about body shape at the height / mass level.

## Splits

The Hugging Face configuration loads `manifest_train.csv` (split `train`, which holds the `split_addb_val` train and val persons, select val with `split_addb_val == "val"`) and `manifest_test.csv` (split `test`).

`split_addb` (force benchmark train / test, person level) with `split_addb_val` (validation persons drawn from the train persons), and `split_t2m` (text-to-motion train / val / test / patho, person level). No person is in two folds of one split.

All three are drawn with a stratified, seeded person-level search (2026-10-09): test and val persons cover the height x body-mass tertile cells of the healthy persons and the bottom / top 10% of height and mass, use different studies within a cell where possible, and cover as many activities as possible, while train keeps every activity, study and body-size cell that any other fold has. Test and val persons are drawn only from persons with at least 90% usable trials (split_addb test and val also need at least 90% trials with `eval_grf_trial_ok`). Pathological persons are `split_t2m=patho` and keep their earlier `split_addb` (separate `eval_stratum`). `split_addb` therefore no longer equals the split shipped with the AddBiomechanics source data. Splits of earlier versions are not compatible: caches or checkpoints trained on an earlier split may contain current test persons.

| split_addb | persons | trials |
|---|---|---|
| test | 19 | 914 |
| train | 121 | 9,472 |

| split_addb_val | persons | trials |
|---|---|---|
| test | 19 | 914 |
| train | 109 | 8,374 |
| val | 12 | 1,098 |

| split_t2m | persons |
|---|---|
| patho | 5 |
| test | 35 |
| train | 83 |
| val | 17 |

The split_t2m walking test set has 4 persons with a body mass target of 85 kg or more. The running test set has 0 (running is not evaluable at high body mass).

## Health status and sensitive information

| health_status | persons | trials |
|---|---|---|
| TKA | 4 | 78 |
| healthy | 135 | 10,243 |
| stroke | 1 | 65 |

The data include persons after total knee arthroplasty and after stroke (`health_status`, `health_detail`). This is health information. `trial_name` keeps the source names: 240 Han2023 names contain a capture date and the Fregly2012 names (trials: 78) contain subject initials. Ages range from 18 to 86 years (unknown in 200 trials). Do not try to re-identify persons.

## Intended and out-of-scope uses

Intended: research on motion-to-force and force-aware motion models, text-to-motion with physical labels, biomechanics method development. Out of scope: clinical decisions or diagnosis, identifying or profiling persons, and human mesh recovery from images or video (there are no images or video in ForceBody).

## Paper numbers

The benchmark numbers of the camera-ready paper (the ForceBody rows of Table 3, and Tables 4, 5, 14 and 15) are the mean and standard deviation over three seeds of the baselines retrained on release 1.2.1. The benchmark inputs, labels, masks and splits of 1.2.1 are identical in 1.3.0. Checkpoints were selected on the 12 validation persons (`split_addb_val == "val"`, 1,098 trials), and the 19 test persons (914 trials, healthy and pathological strata) were scored once. The Benchmark section of the GitHub README gives Tables 3, 4 and 5. The AddB rows of Table 3 are copied from the reviewed build and serve as a reference, not as a matched comparison. They used dynamics-pass (pass 2) kinematics as input (see Leakage policy). The submitted version of the paper used the reviewed build 1.0.0 with the v1 labels. The torque labels moved between versions. Scoring the v1.1 `tau_selfconsistent` against the reviewed-build labels gives RMSE / MAE / PCC 0.287 / 0.087 / 0.721, the same order as the model errors, so numbers from different versions must not be mixed in one table.

## Known issues

- vanderZee2022. Force-plate cross-talk remains after the plate re-zero (swing-foot readings depend on the other plate load). The root residual force is higher than in other studies.
- Carter2023 downhill. AddB did not dynamics-process these kinematics. The root residual stays high and the torque labels are flagged. `grf_valid_mask_plate` is offered for GRF supervision.
- Carter2023 trials whose raw plate CoP is unusable keep a fixed-point CoP (`qc_cop_fixed_point`, `qc_tau_cop_unreliable`). Exclude them from CoP and torque evaluation. Their forces are valid. Since 1.2.1 the loaded frames are in `grf_frame_ok`, while `cop_frame_ok` and `tau_frame_ok` exclude them.
- Han2023 yoga / dance: a few extreme-torque frames of kinematic origin (`tau_frame_ok` excludes them).
- vanderZee2022 S09 has no upper-body markers (`qc_upper_body_unobserved`). AddB pass 0 and pass 2 arms differ by about 200 mm there and the SKEL arms, trunk and head are unconstrained (masked).
- The scapula DOFs are pinned at their bounds in most fits (`qc_skel_scapula_pinned`). Scapula elevation sits on its +-0.1 rad bound (size unsourced, SKEL issue #19) in many kinematics fits. Mask scapula DOFs in losses.
- SKEL left ankle axis (SKEL issue #45, kept upstream): `ankle_angle_l` / `subtalar_angle_l` are not the mirror of the right DOFs and the fit compensates with subtalar_l. Do not mirror motions by swapping L/R SKEL DOFs. Compare left and right ankles with the AddB q (`q_kin`, `q_gt`). Read the poses with official SKEL (a SKEL with the issue #45 fix moves the left toes by about 22 mm at p95).
- Joint-limit bounds of the fit are left/right asymmetric for shoulder_y and pro_sup (unsourced kin_skel values, kept).
- Pelvis joint error is higher than in the first v1.2 fit (median about 2 mm) after the knee weight change. The SKEL femur is shorter than the AddB femur (shape-step limitation).
- Head post-fit. A few trials sit on a neck limit for many frames (`qc_head_at_limit_frac_*`). Trials without head markers keep the head rigid with the thorax (neck line 3-10 deg forward of the markers in head-marker studies).
- Non-arm fit-only spikes (head, spine) exist in a minority of trials. They are masked per joint group in `fit_event_ok_*`, not repaired. Head post-fit steps > 0.5 rad per frame are masked in the head column.
- Dynamics-pass target repair. On repaired trials the Newton consistency of `poses_dyn` with the GRF dropped by more than 0.02 (correlation) in a minority of trials and the lower-body error rose by more than 1 mm in some. Accepted. `qc_skel_vs_qgt_err_mm` compares with `q_gt`, not with the repaired target.
- Some dyn-pass arm events listed by the 2026-10-07 diagnosis refer to the dyn fit before the 2026-10-08/09 target repair. They are kept in `arm_event_ok_dyn` (conservative), together with the events of the re-scan of the final fits.
- Carter2023 running. In some trials AddB marks the flight frames as missing GRF (`grf_valid_mask` False), so force labels and the T-GRF benchmark cover stance frames only there (most trials keep the flight frames valid with zero force).
- `floor_height_m` is the loaded step height on stairs / step activities. No floor normal is shipped.
- `visualize.py` (GitHub repository, `code/`) needs the public SKEL package, the SKEL model files and an EGL-capable renderer (see `code/VISUALIZATION.md`). Reading the data needs only numpy.
- Release verification left small deviations. In one trial the fit joint error recomputed from the stored float32 `joints` / `addb_joints_kin` differs from the stored `qc_skel_fit_err_mm` by 0.054 mm (tolerance 0.05 mm, so one of 10,386 trials fails the independent reproduction check). The median lag between GRF and kinematic toe-off events is -74.5 ms for Uhlrich2023 (all other studies -50 to 0 ms), just outside the -70 to +10 ms window used as a sanity range.
- No acceleration filter was applied to the kinematics-pass target: `poses` keep a small frame-to-frame wobble on trials whose AddB pass-0 inverse kinematics is noisy. The dynamics-pass target is smoother (AddB dynamics pass, repaired targets use the AddB acceleration-minimizing filter).
- Name clash. `chair_contact` in manifest.csv is the trial-level protocol flag (a chair is used in the trial), while `chair_contact` inside the trial file is the frame-level [T] seated mask of the Falisse2017 chair channel (all False without a chair channel). Use `qc_chair_channel_active` for the trial-level summary of the frame mask.
- CoP placement at low load. Frames where a foot channel is loaded (vertical force above 5% BW) and its CoP lies more than 0.15 m from both feet are not all flagged. A scan of the release found 3,227 GRF-valid such frames in 664 trials with `cop_frame_ok` True and `cop_off_foot` False. Mask CoP losses by load if this matters.
- Han2023 `sit` (12 trials) and `sit-to-stand` (2 trials): in the frames a posture rule marks as seated, the foot GRF stays near 1 BW (median 0.999 BW) and there is no seat force channel, so either the seat load is not measured or the label does not describe a loaded seated phase. The labels were not changed.
- Treadmill belt residual. On treadmill trials the stance foot is not perfectly stationary after the belt motion is removed with a kinematic estimate of the horizontal belt speed (and the protocol grade for the vertical component): the median stance-foot speed is about 0.05-0.10 m/s depending on the trial sample and pose set, against about 0.02-0.05 m/s overground.
- SKEL knee clamped at full extension. In many walking trials the SKEL knee sits on its 0 bound while the AddB target knee is flexed about 3-8 deg, more often on the left. Flagged per frame in `knee_clamp_frame_kin` / `_dyn` since 1.2.1 (1,061 trials above 10% of frames in `poses`). Not refitted. On kin_retimed trials `knee_clamp_frame_dyn` compares `poses_dyn` with `q_gt`, not with the retimed kinematic target.
- Stairs (22 trials in activity class stairs). No per-step floor is shipped. `floor_height_m` is one loaded-step height per trial, and one fitted floor plane per trial (as in the T2M encoders) treats a staircase as a ramp, so heights above that floor are not step-relative.
- Moore2015 standing (29 trials). In the wide-stance phases the feet sit 2-4 cm below the trial median floor level on about 40% of frames. Whether this is real penetration or a floor estimate set too high is not resolved.
- Pathological persons (Fregly2012 TKA 78 trials, Li2021 stroke 65 trials). The `caption_*` phrases never mention the pathology, and the Li2021_S05 fits have about 1.8 times the healthy fit error (standing 26.6 mm vs 14.7 mm). Use them as an out-of-distribution stratum (`eval_stratum`).
- Trials at 201 Hz and 250 Hz (by rounded `fps_true_grid`, 24 Han2023 trials and 1 Carter2023 trial at 250 Hz, 111 Tiziana2019 trials at 201 Hz) ship on their native grid. Resampling them to 100 Hz without a low-pass filter (as the T2M encoders do) aliases force-plate noise (Han2023: 7-30 N RMS below 50 Hz, peaks are kept).
- vanderZee2022 walking. The median peak hip flexion torque is 1.95 N m/kg, against 1.0 (Carter2023) and 1.36 (Moore2015). The cause is not identified (possibly the plate cross-talk above).
- `q_kin` beyond 2 pi on a shoulder DOF in a few trials (`qc_q_kin_wrap`): an AddB Euler branch of the pass-0 target, not unwrapped. The arm frames are masked in `arm_event_ok_kin` (1.2.1). The mask covers only the frames beyond 2 pi (+-0.1 s). In some of these trials the wrapped stretch is wider (an arm DOF of `q_kin` more than pi/2 from `q_gt` outside the mask), so skip arm losses on the whole trial if in doubt.
- Constant-labelled treadmill belts that change speed (`qc_belt_speed_change`, `_suspected`) and segments whose label does not match the motion (`label_conflict_codes` SEGMENT_MOTION_MISMATCH) are flagged in 1.2.1. Trials and labels are unchanged.

## License and attribution

The ForceBody data are released under the Creative Commons Attribution-NonCommercial 4.0 International license (CC BY-NC 4.0). Commercial use is not permitted. LICENSE.md holds the full terms.

- Fields derived from AddBiomechanics (forces, AddB kinematics, torques, uncertainty, metadata) build on the AddBiomechanics dataset (CC BY 4.0). Cite AddBiomechanics and every source study you use.
- SKEL-derived data (the trial fields `poses*`, `trans*`, `joints*`, `betas`, `skel_gender`, the persons.csv columns `beta_0` to `beta_9` and `shapy_*`, and marker_vertex_map.json) are also subject to the SKEL model license, which allows non-commercial scientific research only. Building meshes from these fields requires the SKEL model, which you obtain from its authors under their license.
- The code in the GitHub repository (`forcebody_io.py`, `visualize.py`, `benchmark/`, `baselines/` in `code/`) is released under the MIT license. This Hugging Face repository holds no code.

## How to cite

```bibtex
@inproceedings{kwon2026forcebody,
  title     = {ForceBody: Force-Paired Parametric Body Motion with Torque Uncertainty},
  author    = {Kwon, Joonwoo and Zhang, Yufei and Liu, Xiaoming and Cui, Zijun},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Evaluations and Datasets Track},
  year      = {2026}
}
@inproceedings{werling2024addbiomechanics,
  title     = {AddBiomechanics Dataset: Capturing the Physics of Human Motion at Scale},
  author    = {Werling, Keenon and Kaneda, Janelle and Tan, Tian and Agarwal, Rishi and Skov, Six and Van Wouwe, Tom and Uhlrich, Scott and Bianco, Nicholas and Ong, Carmichael and Falisse, Antoine and Sapkota, Shardul and Chandra, Aidan and Carter, Joshua and Preatoni, Ezio and Fregly, Benjamin and Hicks, Jennifer and Delp, Scott and Liu, C. Karen},
  booktitle = {European Conference on Computer Vision (ECCV 2024)},
  series    = {Lecture Notes in Computer Science},
  pages     = {490--508},
  publisher = {Springer},
  year      = {2024},
  doi       = {10.1007/978-3-031-73223-2_27}
}
@article{werling2023addbiomechanics,
  title   = {AddBiomechanics: Automating model scaling, inverse kinematics, and inverse dynamics from human motion data through sequential optimization},
  author  = {Werling, Keenon and Bianco, Nicholas A. and Raitor, Michael and Stingel, Jon and Hicks, Jennifer L. and Collins, Steven H. and Delp, Scott L. and Liu, C. Karen},
  journal = {PLOS ONE},
  volume  = {18},
  number  = {11},
  pages   = {e0295152},
  year    = {2023},
  doi     = {10.1371/journal.pone.0295152}
}

@article{keller2023skin,
  title   = {From Skin to Skeleton: Towards Biomechanically Accurate {3D} Digital Humans},
  author  = {Keller, Marilyn and Werling, Keenon and Shin, Soyong and Delp, Scott and Pujades, Sergi and Liu, C. Karen and Black, Michael J.},
  journal = {ACM Transactions on Graphics},
  volume  = {42},
  number  = {6},
  pages   = {1--12},
  year    = {2023},
  doi     = {10.1145/3618381}
}
```

Source studies. Cite every study whose trials you use (manifest columns `study` and `study_citation`). AddBiomechanics data: https://addbiomechanics.org/.

| `study` | `study_citation` | reference |
|---|---|---|
| Carter2023 | Carter2023 | Carter, J., Chen, X., Cazzola, D., Trewartha, G., Preatoni, E. Full body kinematics and ground reaction forces of fifty heterogeneous runners completing treadmill running at various speeds and gradients. University of Bath Research Data Archive, 2024. https://doi.org/10.15125/BATH-01341 |
| Falisse2017 | Falisse2017 | Falisse, A., Van Rossom, S., Jonkers, I., De Groote, F. EMG-driven optimal estimation of subject-specific Hill model muscle-tendon parameters of the knee joint actuators. IEEE Transactions on Biomedical Engineering 64(9), 2253-2262, 2017. https://doi.org/10.1109/TBME.2016.2630009 |
| Fregly2012 | Fregly2012 | Fregly, B.J., Besier, T.F., Lloyd, D.G., Delp, S.L., Banks, S.A., Pandy, M.G., D'Lima, D.D. Grand challenge competition to predict in vivo knee loads. Journal of Orthopaedic Research 30(4), 503-513, 2012. https://doi.org/10.1002/jor.22023 |
| Hamner2013 | Hamner2013 | Hamner, S.R., Delp, S.L. Muscle contributions to fore-aft and vertical body mass center accelerations over a range of running speeds. Journal of Biomechanics 46(4), 780-787, 2013. https://doi.org/10.1016/j.jbiomech.2012.11.024 |
| Han2023 | Han2023 | Han, X., Senderling, B., To, S., Kumar, D., Whiting, E., Saito, J. GroundLink: A dataset unifying human body movement and ground reaction dynamics. SIGGRAPH Asia 2023 Conference Papers, 1-10, 2023. https://doi.org/10.1145/3610548.3618247 |
| Li2021 | Li2021 | Li, G., Shourijeh, M.S., Ao, D., Patten, C., Fregly, B.J. How well do commonly used co-contraction indices approximate lower limb joint stiffness trends during gait for individuals post-stroke? Frontiers in Bioengineering and Biotechnology 8, 588908, 2021. https://doi.org/10.3389/fbioe.2020.588908 |
| Moore2015 | Moore2015 | Moore, J.K., Hnat, S.K., van den Bogert, A.J. An elaborate data set on human gait and the effect of mechanical perturbations. PeerJ 3, e918, 2015. https://doi.org/10.7717/peerj.918 |
| Tiziana2019 | Lencioni2019 | Lencioni, T., Carpinella, I., Rabuffetti, M., Marzegan, A., Ferrarin, M. Human kinematic, kinetic and EMG data during different walking and stair ascending and descending tasks. Scientific Data 6, 309, 2019. https://doi.org/10.1038/s41597-019-0323-z |
| Uhlrich2023 | Uhlrich2023 | Uhlrich, S.D., Falisse, A., Kidziński, Ł., Muccini, J., Ko, M., Chaudhari, A.S., Hicks, J.L., Delp, S.L. OpenCap: Human movement dynamics from smartphone videos. PLOS Computational Biology 19(10), e1011462, 2023. https://doi.org/10.1371/journal.pcbi.1011462 |
| vanderZee2022 | vanderZee2022 | van der Zee, T.J., Mundinger, E.M., Kuo, A.D. A biomechanics dataset of healthy human walking at various speeds, step lengths and step widths. Scientific Data 9, 704, 2022. https://doi.org/10.1038/s41597-022-01817-1 |

## Versions

| version | date | note |
|---|---|---|
| 1.0.0-reviewed | 2026-05-07 | reviewed build (numbers of the submitted paper) |
| internal v1 | 2026-08-31 | internal build |
| internal 1.1 | 2026-09-27 | internal build (frame and timebase fixes) |
| 1.2.0 | 2026-10-11 | first frozen 1.2 build (first build 2026-10-09). Superseded by 1.2.1 (QC mask erratum) |
| 1.2.1 | 2026-10-12 | 1.2.0 with corrected QC masks and flags (pose, force and torque arrays unchanged) |
| 1.3.0 | 2026-10-29 | this release: Hamner2013 study id (40 trial keys renamed, map shipped), `sex_resolved` corrected, benchmark results, documentation. Arrays otherwise identical to 1.2.1, see CHANGELOG.md. Documentation updated on 2026-11-01 and 2026-11-02 (SCHEMA_v1.3.csv descriptions only, no other data file changed) |

## Maintenance

Report errors and questions on the GitHub issues page https://github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026/issues. CHANGELOG.md tracks issues and corrections per version. Replaced arrays are kept in an internal v1.1 provenance archive (not uploaded), so every change can be traced back to v1.1.