ForceBody: Force-Paired Parametric Body Motion with Torque Uncertainty

Joonwoo Kwon, Yufei Zhang, Xiaoming Liu, Zijun CuiarXiv 2026 · arXiv

TL;DR

ForceBody is the first large-scale dataset to pair parametric body motion (using the biomechanically grounded SKEL model) with measured ground reaction forces and joint torques. Crucially, it is the first to provide per-frame, per-joint uncertainty estimates for these torques, enabling more robust, physics-aware learning for human motion reconstruction and generation.

What Problem Does This Paper Address?

In the worlds of computer vision and graphics, we’ve become very good at reconstructing 3D human poses using parametric models like SMPL. However, these models are often "hollow"—they describe the surface but ignore the underlying forces and torques that actually drive movement.

Biomechanics researchers have the opposite problem: they have rich data on ground reaction forces (GRF) and joint torques, but they typically represent motion using sparse marker trajectories or specialized skeletons that don't play well with modern deep learning pipelines. Furthermore, joint torques aren't measured directly; they are calculated via inverse dynamics, a process notoriously sensitive to even tiny amounts of noise in the input motion data. Until now, no dataset quantified how much we should trust these derived torque labels.

ForceBody bridges this gap by aligning biomechanical force measurements with a parametric body model and providing a "reliability score" (uncertainty) for every torque annotation.

Methods

The authors developed a sophisticated pipeline to transform raw biomechanical data into a format ready for machine learning:

  1. SKEL Motion Fitting: They adopted SKEL, a model that places a biomechanically accurate skeleton inside an SMPL-compatible surface. They used a two-stage optimization process to fit SKEL to the marker data from the AddBiomechanics dataset, ensuring the resulting motion was both visually accurate and physically plausible. SKEL adopts a biomechanical skeleton (left) inside the SMPL surface (right).

  2. Uncertainty Quantification: To address the sensitivity of inverse dynamics, the team used Monte Carlo sampling. For every frame, they perturbed the joint angles 1,000 times with realistic noise and ran these samples through a physics solver. The resulting spread (standard deviation) of torques defines the uncertainty ($\sigma_\tau$) for that frame. Monte Carlo torque distribution for a single frame. The spread represents the released uncertainty.

  3. Benchmarking: They evaluated six different neural architectures (including LSTMs, Transformers, and the newer Mamba2) on the task of predicting forces and torques directly from motion.

SKEL adopts a biomechanical skeleton (left) inside the SMPL surface (right).
Monte Carlo torque distribution for a single frame. The spread represents the released uncertainty.

Datasets

The primary contribution is the ForceBody dataset:

  • Scale: 10,386 motion trials, 9.7 million frames, and nearly 27 hours of data.
  • Diversity: Data from 140 subjects performing activities like walking, running, jumping, and stair climbing.
  • Comparison: Unlike previous datasets like GroundLink (small scale, no torques) or ImDy (simulated forces), ForceBody provides measured forces and derived torques on a parametric model at scale.

Key Insights and Ideas

  • Uncertainty is Motion-Dependent: One of the most interesting findings is that relative uncertainty is lower during dynamic movements (like jumping) and higher during near-static ones (like standing). In static poses, the actual torque signal is small, so the "noise" from kinematic errors becomes a much larger fraction of the total signal.
  • Lever Arms Matter: Uncertainty is not uniform across the body. Proximal joints like the hip and knee show much higher absolute uncertainty because they carry more body weight and have longer lever arms, making them more sensitive to angular errors. Bones colored by torque magnitude (warm) vs. uncertainty (cool).

  • Physics-Grounded Supervision: By using SKEL, the dataset ensures that joint torques are applied to anatomically correct axes, which is a prerequisite for valid biomechanical analysis that SMPL's non-anatomical joints cannot provide.

Bones colored by torque magnitude (warm) vs. uncertainty (cool).

Key Results

  • High Fidelity: The SKEL fitting process achieved a mean per-joint error of only 21.6 mm, and the angular accelerations of the fitted model matched the original data with a Pearson correlation of 0.97 to 1.00. Comparison of angular accelerations between the original data and the ForceBody SKEL fit.

  • Model Performance: Among the tested architectures, Mamba2 emerged as the top performer for joint torque prediction.

  • The Value of Uncertainty: The authors demonstrated that the uncertainty labels aren't just for show. By finetuning a Transformer using uncertainty-weighted loss (essentially telling the model to care less about samples with high noise), they reduced the Mean Absolute Error (MAE) of torque estimation by 5%.
Comparison of angular accelerations between the original data and the ForceBody SKEL fit.

Limitations and Future Work

  • Data Imbalance: The dataset is heavily skewed toward walking and running (over 93% of trials). Rare activities like stair climbing or squats have much less representation.
  • Simplified Noise Model: The current Monte Carlo approach only accounts for noise in joint angles. Future iterations could incorporate errors from ground reaction force measurements, center of pressure estimates, and body segment inertia.
  • Shape Space: The fitting is limited by the representational capacity of the SKEL shape space, which may struggle with atypical body types.

ForceBody represents a significant step toward "physics-complete" human digital twins, providing the community with the data needed to move beyond just how humans look when they move, to how they actually move.

Summary updated

Abstract

Forces and torques drive human motion, central to biomechanics, robotics, and physics-aware animation, yet they are often absent from learning-based motion pipelines for body reconstruction and generation. Some motion datasets provide force annotations, but the paired motion sequences are incompatible with the parametric body models used in modern learning pipelines. In addition, existing datasets typically obtain joint torques through inverse dynamics, which is inherently sensitive to the quality of the input kinematics; however, they do not quantify the reliability of these torque estimates. We close this gap with ForceBody, the first dataset to bring ground truth force supervision into a parametric body representation, shipped with per-sample torque uncertainty. ForceBody pairs the SKEL body model with ground reaction forces and inverse-dynamics joint torques across 10,386 motion trials (9.7M frames, 26.9 hours) from 140 subjects. Each torque label is paired with a per-frame, per-joint uncertainty annotation obtained by Monte Carlo sampling through the inverse dynamics pipeline. We benchmark six architectures (MLP, Conv1D, LSTM, GRU, Mamba2, Transformer) on ground reaction force and joint torque prediction from SKEL motion. As one example use of the released uncertainty, finetuning a Transformer with per-sample uncertainty weighting reduces the MAE of joint torque estimation by 5%. Code and data are at https://github.com/DominoAI-Lab/ForceBody-NeurIPS-ED-2026.
BibTeX
@misc{kwon2026forcebody,
  author = {Joonwoo Kwon and Yufei Zhang and Xiaoming Liu and Zijun Cui},
  title = {{ForceBody: Force-Paired Parametric Body Motion with Torque Uncertainty}},
  year = {2026},
  eprint = {2611.22769},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22769},
}
Figure