Datasets:

kwonjoon
/

ForceBody

---
license: cc-by-nc-4.0
language:
- en
pretty_name: ForceBody
size_categories:
- 10K<n<100K
task_categories:
- other
tags:
- 3d-human-body-models
- biomechanics
- ground-reaction-force
- joint-torque
- skel
- human-motion
configs:
- config_name: default
  data_files:
  - split: train
    path: manifest_train.csv
  - split: test
    path: manifest_test.csv
---
<div align="center">

<h1>ForceBody: Force-Paired Parametric Body Motion with Torque Uncertainty</h1>

**NeurIPS 2026, Evaluations & Datasets Track**

Joonwoo Kwon<sup>1</sup>, Yufei Zhang<sup>2</sup>, Xiaoming Liu<sup>3</sup>, Zijun Cui<sup>1</sup>

<sup>1</sup>Michigan State University, <sup>2</sup>Independent Researcher, <sup>3</sup>University of North Carolina at Chapel Hill

<img src="assets/divider.png" width="100%" height="4">

### 💻 Code and benchmark on GitHub: [github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026](https://github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026)

</div>

<p align="center"><img src="assets/teaser.png" width="100%"></p>

ForceBody pairs SKEL body motion with measured ground reaction forces (GRF), inverse-dynamics joint torques, and a per-frame, per-joint torque uncertainty σ<sub>τ</sub>. This page holds the data. Code, loader and benchmark are on GitHub. The source data come from [AddBiomechanics](https://addbiomechanics.org/). The Croissant metadata, with Responsible AI fields, is in [croissant.json](https://huggingface.co/datasets/kwonjoon/ForceBody/resolve/main/croissant.json).

| Trials | Subjects | Hours | Frames | Sources | Release |
|:---:|:---:|:---:|:---:|:---:|:---:|
| 10,386 | 140 | 27 | 9.7M | 10 studies (AddBiomechanics) | 1.3.0 (2026-10-29) |

<img src="assets/divider.png" width="100%" height="4">

## Demos

| Running | Stair ascent |
|:---:|:---:|
| <img src="assets/running_hamner2013.gif" width="100%"> | <img src="assets/stair_ascent_falisse2017.gif" width="100%"> |
| **Sit-to-stand** | **Drop jump** |
| <img src="assets/sit_to_stand_uhlrich2023.gif" width="100%"> | <img src="assets/drop_jump.gif" width="100%"> |

**Torque |τ| (left body) and uncertainty σ<sub>τ</sub> (right body)**

<p align="center"><img src="assets/activities_grid_tau_sigma.gif" width="100%"></p>

<img src="assets/divider.png" width="100%" height="4">

## Quick start

### 0. Download the SKEL model (required)

The trial files store SKEL parameters (`poses`, `betas`, `trans`), not meshes. Register and download the SKEL model from the [SKEL website](https://skel.is.tue.mpg.de/) under its non-commercial license. ForceBody does not redistribute it. You need it only to build meshes. The joint positions (`joints`), force, torque and uncertainty arrays can be read without it.

### 1. Download the data

```bash
pip install -U huggingface_hub
hf download kwonjoon/ForceBody --repo-type dataset --local-dir ForceBody
```

The full dataset is about 46 GB (43 GiB). Keep about 50 GB of disk free. You do not need a Hugging Face login, and an `HF_TOKEN` only raises the rate limit.

To try one trial first (9 MB), run this command instead of the `hf download` line above. Steps 2 and 3 then run unchanged.

```bash
hf download kwonjoon/ForceBody trials/Moore2015/S10/Moore2015_S10_walk_014.npz --repo-type dataset --local-dir ForceBody
```

### 2. Install the loader

Run these commands in the folder where you ran step 1, so that `ForceBody/` and `ForceBody-NeurIPS-ED-2026/` are side by side.

```bash
git clone https://github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026.git
cd ForceBody-NeurIPS-ED-2026/code
pip install numpy
```

### 3. Load a trial

Run Python in `ForceBody-NeurIPS-ED-2026/code`. The data folder of step 1 is then `../../ForceBody`.

```python
import forcebody_io as fio
d = fio.load_trial("../../ForceBody/trials/Moore2015/S10/Moore2015_S10_walk_014.npz", pose_source="pass0")
m = float(d["mass_kg_dynamics"])
ok = d["mc_frame_ok"] & bool(d["torque_label_usable"])   # frame mask and trial flag
tau   = d["tau_selfconsistent"][ok, 6:] / m   # joint torques (N m / kg), columns 0-5 are root residuals
sigma = d["sigma_tau"][ok, 6:] / m            # torque uncertainty (N m / kg)
print(tau.shape, sigma.shape)
```

For force or torque prediction, feed the model `pose_source="pass0"`. Combine every frame mask with its trial flag.

### 4. Train and evaluate

See the GitHub repository (`code/baselines/`) for cache building, training, uncertainty-weighted finetuning, inference and scoring.

<img src="assets/divider.png" width="100%" height="4">

## What is in a trial

T = frames. C = contact channels: 2 (right foot, left foot), or 3 in Falisse2017, where channel 2 is the pelvis on an instrumented chair.

| Field | Shape | Content |
|---|---|---|
| `poses`, `trans`, `betas` | `[T,46]`, `[T,3]`, `[10]` | SKEL pose, root translation, body shape |
| `grf_aligned`, `cop_aligned_foot` | `[T,C,3]` | Measured GRF (N) and centre of pressure (m) per contact channel |
| `tau_selfconsistent` | `[T,37]` | Inverse dynamics in AddBiomechanics DOF order (`dof_names`). Columns 0-5 are pelvis root residuals (0-2 moments in N m, 3-5 forces in N). Columns 6-36 are joint torques (N m) |
| `sigma_tau`, `mu_tau` | `[T,37]` | Monte Carlo torque uncertainty and mean, same columns |
| `grf_frame_ok`, `tau_frame_ok`, `mc_frame_ok` | `[T]` | Frame masks for GRF, torque and σ<sub>τ</sub> |
| `cop_frame_ok`, `grf_foot_ok` | `[T,2]` | Per-foot frame masks (right, left) |
| `force_label_usable`, `torque_label_usable` | scalar | Trial flags |

Three σ<sub>τ</sub> counts appear in the release. The trial files carry σ<sub>τ</sub> for 10,382 of the 10,386 trials (`has_uncertainty`). The other 4 trials are shorter than one period of the 6 Hz noise band. The σ<sub>τ</sub> statistics of the paper (Sec. 3.2, Table 2, Appendix D) use the 8,732 trials with `has_uncertainty` and GRF valid on at least 50% of frames (`grf_valid_frac` >= 0.5). The paper calls them the trials with released σ<sub>τ</sub>. The `subjects.csv` column `n_with_uncertainty` holds the v1.1 count (8,652 trials in total). Use `has_uncertainty` in `manifest.csv` instead.

<img src="assets/divider.png" width="100%" height="4">

## Splits

The splits are at subject level (`split_addb`). The Hugging Face split `train` (`manifest_train.csv`) holds 121 subjects and 9,472 trials. It includes the 12 validation subjects (1,098 trials, `split_addb_val == "val"`) that the paper uses for checkpoint selection. The split `test` (`manifest_test.csv`) holds 19 subjects and 914 trials. Select checkpoints on the validation subjects only. DETAILS.md ("Splits") gives the split rules and the text-to-motion split `split_t2m`.

<img src="assets/divider.png" width="100%" height="4">

## Files

| Path | Content |
|---|---|
| `trials/<study>/S<NN>/<trial_uid>.npz` | One file per trial |
| `models/` | AddBiomechanics OpenSim models (`.osim.gz`) per source b3d file, processing passes 0 and 2 (the torques use the pass-2 model) |
| `manifest.csv` (+ `_train`, `_test`) | One row per trial: metadata, split, quality flags |
| `persons.csv` | One row per subject: SKEL shape (`beta_0` to `beta_9`), sex, height, mass, body measurements, splits |
| `subjects.csv` | One row per subject: age, health status, footwear, mass sources, trial counts, activities (`n_with_uncertainty` is the v1.1 count, see DETAILS.md) |
| `SCHEMA_v1.3.csv` | Dtype, shape, unit, frame and description of all 547 trial fields and of every `manifest.csv` and `persons.csv` column |
| `DETAILS.md` | Full documentation (masks, conventions, known issues) |
| `CHANGELOG.md`, `MD5SUMS.txt` | Version history, md5 checksums |
| [`croissant.json`](https://huggingface.co/datasets/kwonjoon/ForceBody/resolve/main/croissant.json) | Croissant metadata with Responsible AI fields |
| `LICENSE.md` | License and attribution terms |
| Other files | Gait events, key maps and marker map (DETAILS.md, "Layout") |

<img src="assets/divider.png" width="100%" height="4">

## License

Data under **CC BY-NC 4.0** (non-commercial). SKEL-derived fields also follow the [SKEL license](https://skel.is.tue.mpg.de/). Code on GitHub under MIT. Cite [AddBiomechanics](https://addbiomechanics.org/) and the source studies you use. DETAILS.md ("How to cite") gives the BibTeX entries and the source-study references. LICENSE.md gives the full license and attribution terms.

<img src="assets/divider.png" width="100%" height="4">

## Issues and contact

Please report errors and questions on the GitHub issues page [github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026/issues](https://github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026/issues). CHANGELOG.md records every correction.

<img src="assets/divider.png" width="100%" height="4">

## Citation

```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}
}
```

	---
	license: cc-by-nc-4.0
	language:
	- en
	pretty_name: ForceBody
	size_categories:
	- 10K<n<100K
	task_categories:
	- other
	tags:
	- 3d-human-body-models
	- biomechanics
	- ground-reaction-force
	- joint-torque
	- skel
	- human-motion
	configs:
	- config_name: default
	data_files:
	- split: train
	path: manifest_train.csv
	- split: test
	path: manifest_test.csv
	---

	<div align="center">

	<h1>ForceBody: Force-Paired Parametric Body Motion with Torque Uncertainty</h1>

	NeurIPS 2026, Evaluations & Datasets Track

	Joonwoo Kwon<sup>1</sup>, Yufei Zhang<sup>2</sup>, Xiaoming Liu<sup>3</sup>, Zijun Cui<sup>1</sup>

	<sup>1</sup>Michigan State University, <sup>2</sup>Independent Researcher, <sup>3</sup>University of North Carolina at Chapel Hill

	<img src="assets/divider.png" width="100%" height="4">

	### 💻 Code and benchmark on GitHub: [github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026](https://github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026)

	</div>

	<p align="center"><img src="assets/teaser.png" width="100%"></p>

	ForceBody pairs SKEL body motion with measured ground reaction forces (GRF), inverse-dynamics joint torques, and a per-frame, per-joint torque uncertainty σ<sub>τ</sub>. This page holds the data. Code, loader and benchmark are on GitHub. The source data come from [AddBiomechanics](https://addbiomechanics.org/). The Croissant metadata, with Responsible AI fields, is in [croissant.json](https://huggingface.co/datasets/kwonjoon/ForceBody/resolve/main/croissant.json).

	\| Trials \| Subjects \| Hours \| Frames \| Sources \| Release \|
	\|:---:\|:---:\|:---:\|:---:\|:---:\|:---:\|
	\| 10,386 \| 140 \| 27 \| 9.7M \| 10 studies (AddBiomechanics) \| 1.3.0 (2026-10-29) \|

	<img src="assets/divider.png" width="100%" height="4">

	## Demos

	\| Running \| Stair ascent \|
	\|:---:\|:---:\|
	\| <img src="assets/running_hamner2013.gif" width="100%"> \| <img src="assets/stair_ascent_falisse2017.gif" width="100%"> \|
	\| Sit-to-stand \| Drop jump \|
	\| <img src="assets/sit_to_stand_uhlrich2023.gif" width="100%"> \| <img src="assets/drop_jump.gif" width="100%"> \|

	Torque \|τ\| (left body) and uncertainty σ<sub>τ</sub> (right body)

	<p align="center"><img src="assets/activities_grid_tau_sigma.gif" width="100%"></p>

	<img src="assets/divider.png" width="100%" height="4">

	## Quick start

	### 0. Download the SKEL model (required)

	The trial files store SKEL parameters (`poses`, `betas`, `trans`), not meshes. Register and download the SKEL model from the [SKEL website](https://skel.is.tue.mpg.de/) under its non-commercial license. ForceBody does not redistribute it. You need it only to build meshes. The joint positions (`joints`), force, torque and uncertainty arrays can be read without it.

	### 1. Download the data

	```bash
	pip install -U huggingface_hub
	hf download kwonjoon/ForceBody --repo-type dataset --local-dir ForceBody
	```

	The full dataset is about 46 GB (43 GiB). Keep about 50 GB of disk free. You do not need a Hugging Face login, and an `HF_TOKEN` only raises the rate limit.

	To try one trial first (9 MB), run this command instead of the `hf download` line above. Steps 2 and 3 then run unchanged.

	```bash
	hf download kwonjoon/ForceBody trials/Moore2015/S10/Moore2015_S10_walk_014.npz --repo-type dataset --local-dir ForceBody
	```

	### 2. Install the loader

	Run these commands in the folder where you ran step 1, so that `ForceBody/` and `ForceBody-NeurIPS-ED-2026/` are side by side.

	```bash
	git clone https://github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026.git
	cd ForceBody-NeurIPS-ED-2026/code
	pip install numpy
	```

	### 3. Load a trial

	Run Python in `ForceBody-NeurIPS-ED-2026/code`. The data folder of step 1 is then `../../ForceBody`.

	```python
	import forcebody_io as fio

	d = fio.load_trial("../../ForceBody/trials/Moore2015/S10/Moore2015_S10_walk_014.npz", pose_source="pass0")
	m = float(d["mass_kg_dynamics"])
	ok = d["mc_frame_ok"] & bool(d["torque_label_usable"]) # frame mask and trial flag
	tau = d["tau_selfconsistent"][ok, 6:] / m # joint torques (N m / kg), columns 0-5 are root residuals
	sigma = d["sigma_tau"][ok, 6:] / m # torque uncertainty (N m / kg)
	print(tau.shape, sigma.shape)
	```

	For force or torque prediction, feed the model `pose_source="pass0"`. Combine every frame mask with its trial flag.

	### 4. Train and evaluate

	See the GitHub repository (`code/baselines/`) for cache building, training, uncertainty-weighted finetuning, inference and scoring.

	<img src="assets/divider.png" width="100%" height="4">

	## What is in a trial

	T = frames. C = contact channels: 2 (right foot, left foot), or 3 in Falisse2017, where channel 2 is the pelvis on an instrumented chair.

	\| Field \| Shape \| Content \|
	\|---\|---\|---\|
	\| `poses`, `trans`, `betas` \| `[T,46]`, `[T,3]`, `[10]` \| SKEL pose, root translation, body shape \|
	\| `grf_aligned`, `cop_aligned_foot` \| `[T,C,3]` \| Measured GRF (N) and centre of pressure (m) per contact channel \|
	\| `tau_selfconsistent` \| `[T,37]` \| Inverse dynamics in AddBiomechanics DOF order (`dof_names`). Columns 0-5 are pelvis root residuals (0-2 moments in N m, 3-5 forces in N). Columns 6-36 are joint torques (N m) \|
	\| `sigma_tau`, `mu_tau` \| `[T,37]` \| Monte Carlo torque uncertainty and mean, same columns \|
	\| `grf_frame_ok`, `tau_frame_ok`, `mc_frame_ok` \| `[T]` \| Frame masks for GRF, torque and σ<sub>τ</sub> \|
	\| `cop_frame_ok`, `grf_foot_ok` \| `[T,2]` \| Per-foot frame masks (right, left) \|
	\| `force_label_usable`, `torque_label_usable` \| scalar \| Trial flags \|

	Three σ<sub>τ</sub> counts appear in the release. The trial files carry σ<sub>τ</sub> for 10,382 of the 10,386 trials (`has_uncertainty`). The other 4 trials are shorter than one period of the 6 Hz noise band. The σ<sub>τ</sub> statistics of the paper (Sec. 3.2, Table 2, Appendix D) use the 8,732 trials with `has_uncertainty` and GRF valid on at least 50% of frames (`grf_valid_frac` >= 0.5). The paper calls them the trials with released σ<sub>τ</sub>. The `subjects.csv` column `n_with_uncertainty` holds the v1.1 count (8,652 trials in total). Use `has_uncertainty` in `manifest.csv` instead.

	<img src="assets/divider.png" width="100%" height="4">

	## Splits

	The splits are at subject level (`split_addb`). The Hugging Face split `train` (`manifest_train.csv`) holds 121 subjects and 9,472 trials. It includes the 12 validation subjects (1,098 trials, `split_addb_val == "val"`) that the paper uses for checkpoint selection. The split `test` (`manifest_test.csv`) holds 19 subjects and 914 trials. Select checkpoints on the validation subjects only. DETAILS.md ("Splits") gives the split rules and the text-to-motion split `split_t2m`.

	<img src="assets/divider.png" width="100%" height="4">

	## Files

	\| Path \| Content \|
	\|---\|---\|
	\| `trials/<study>/S<NN>/<trial_uid>.npz` \| One file per trial \|
	\| `models/` \| AddBiomechanics OpenSim models (`.osim.gz`) per source b3d file, processing passes 0 and 2 (the torques use the pass-2 model) \|
	\| `manifest.csv` (+ `_train`, `_test`) \| One row per trial: metadata, split, quality flags \|
	\| `persons.csv` \| One row per subject: SKEL shape (`beta_0` to `beta_9`), sex, height, mass, body measurements, splits \|
	\| `subjects.csv` \| One row per subject: age, health status, footwear, mass sources, trial counts, activities (`n_with_uncertainty` is the v1.1 count, see DETAILS.md) \|
	\| `SCHEMA_v1.3.csv` \| Dtype, shape, unit, frame and description of all 547 trial fields and of every `manifest.csv` and `persons.csv` column \|
	\| `DETAILS.md` \| Full documentation (masks, conventions, known issues) \|
	\| `CHANGELOG.md`, `MD5SUMS.txt` \| Version history, md5 checksums \|
	\| [`croissant.json`](https://huggingface.co/datasets/kwonjoon/ForceBody/resolve/main/croissant.json) \| Croissant metadata with Responsible AI fields \|
	\| `LICENSE.md` \| License and attribution terms \|
	\| Other files \| Gait events, key maps and marker map (DETAILS.md, "Layout") \|

	<img src="assets/divider.png" width="100%" height="4">

	## License

	Data under CC BY-NC 4.0 (non-commercial). SKEL-derived fields also follow the [SKEL license](https://skel.is.tue.mpg.de/). Code on GitHub under MIT. Cite [AddBiomechanics](https://addbiomechanics.org/) and the source studies you use. DETAILS.md ("How to cite") gives the BibTeX entries and the source-study references. LICENSE.md gives the full license and attribution terms.

	<img src="assets/divider.png" width="100%" height="4">

	## Issues and contact

	Please report errors and questions on the GitHub issues page [github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026/issues](https://github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026/issues). CHANGELOG.md records every correction.

	<img src="assets/divider.png" width="100%" height="4">

	## Citation

	```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}
	}
	```