PhysHanDI: Physics-Based Reconstruction of Hand-Deformable Object Interactions

Jihyun Lee, Changmin Lee, Donghwan Kim, Tae-Kyun Kim

Reconstructing how humans interact with objects is a cornerstone of computer vision, with critical applications ranging from AR/VR to robotic imitation learning. However, the field has historically struggled with objects that bend, fold, or squish.

Existing state-of-the-art approaches typically fall into two traps. First is a rigid bias: models assume objects are entirely rigid or made of a few rigid parts, failing to capture the complex dynamics of a folding towel or a squeezed stuffed animal. Second is sparse interaction: methods that attempt to handle deformation often represent the interacting hand as a sparse set of points (around 30) sampled from depth maps. Because depth sensors suffer from severe occlusions during contact, these sparse points provide a noisy, incomplete picture of the forces at play, leading to unstable simulations and unrealistic object behavior.

To bridge this gap, researchers have introduced PhysHanDI, a framework that achieves dense 3D reconstruction of both hands and non-rigid objects from sparse-view RGB-D video. By treating the hand and the deformable object as a unified, physically coupled system, PhysHanDI produces interactions that are both visually accurate and physically plausible.

The PhysHanDI Approach

PhysHanDI utilizes a three-stage optimization pipeline that tightly couples kinematic tracking with physical simulation:

  1. Dense Hand Reconstruction: The system first fits a MANO model (a standard parametric 3D hand model) to sparse-view RGB-D data. This generates a dense, 778-vertex mesh of the hand, accurately estimating geometry even in areas occluded from the camera's view.
  2. Physics-Based Object Reconstruction: The deformable object is modeled as a Spring-Mass system. Instead of relying on sparse depth points, PhysHanDI creates "virtual springs" between the dense MANO hand vertices and the object's nodes. The hand's motion acts as a boundary condition, physically driving the object's deformation.
  3. Hand Refinement via Inverse Physics: In its most novel step, the framework uses inverse physics to fine-tune the hand's pose. If the initial hand reconstruction produces a simulated object deformation that deviates from the actual video observations, the error is backpropagated through the physics engine to adjust the hand parameters. Essentially, the object's physical deformation acts as a prior to correct and improve hand tracking.

Datasets: Introducing DENSEHDI

To evaluate the framework, the researchers utilized a dense subset of the existing PhysTwin benchmark. More importantly, they introduced DENSEHDI, a new dataset consisting of 19 sequences across 10 different deformable object types (e.g., pouches, towels, hats). Unlike previous datasets that primarily feature "pointy" interactions like two-finger pinching, DENSEHDI focuses on palm-heavy, dense contacts such as wiping, wrapping, or folding, which are notoriously difficult to track.

Key Insights and Results

PhysHanDI demonstrates a "virtuous cycle" in human-object interaction modeling: accurate hand tracking leads to better physics simulations, and the constraints of a valid physics simulation can correct errors in hand tracking.

The framework significantly outperforms the previous state-of-the-art across several key metrics:

  • Density Prevents Force Diffusion: Using a dense hand mesh (778 vertices) rather than sparse points allows for precise localization of contact forces. This prevents the "force diffusion" seen in previous methods, where forces erroneously spread over non-contact regions.
  • Superior Reconstruction and Prediction: On the PhysTwin-dense dataset, PhysHanDI reduced the Chamfer Distance (a measure of 3D shape error) in dynamic regions from 10.78mm to 8.32mm. When predicting how an object will move in unseen future frames, it maintained a higher accuracy (14.35mm CD) compared to baselines (16.32mm+).
  • Optimized Topology: The authors introduced a metric called Radius-to-Resolution Deviation (RRD) to measure simulation stability. By leveraging dense hand meshes, PhysHanDI can use a smaller "connection radius" for its virtual springs, resulting in an RRD 7x lower than baselines. This mathematically sound topology prevents unrealistic jitter and excessive wave dispersion.
  • Robustness to Noise: When 1mm of artificial noise was added to the depth input, PhysHanDI's error increased only marginally, whereas baseline methods suffered significant performance degradation.

Limitations and Future Work

While PhysHanDI represents a major step forward for modeling non-rigid interactions, the authors note a few areas for future improvement. Currently, the model assumes a static contact topology, meaning it does not yet handle sliding or re-grasping where contact points change dynamically. Additionally, the system uses boundary conditions to approximate force rather than calculating explicit pressure (Newtons), which would be necessary for advanced haptic applications. Finally, while inference is fast (0.11s per frame), the inverse physics refinement stage adds computational overhead during the training and optimization process.

Despite these limitations, PhysHanDI successfully proves that coupling dense kinematic hand models with physics-based simulation is a highly effective path forward for understanding complex, real-world human-object interactions.

Summary updated

Abstract

While existing methods for reconstructing hand–object interactions have made impressive progress, they either focus on rigid or part-wise rigid objects—limiting their ability to model real-world objects (e.g., cloth, stuffed animals) that exhibit highly non-rigid deformations—or model deformable objects without full 3D hand reconstruction. To bridge this gap, we present PhysHanDI (Physics-based Reconstruction of Hand and Deformable Object Interactions), a framework that enables full 3D reconstruction of both interacting hands and non-rigid objects. Our key idea is to physically simulate object deformations driven by forces induced from densely reconstructed 3D hand motions, ensuring that the reconstructed object dynamics are both physically plausible and coherent with the interacting hand movements. Furthermore, we demonstrate that such simulation of object deformations can, in turn, refine and improve hand reconstruction via inverse physics. In experiments, PhysHanDI outperforms the state-of-the-art baseline across reconstruction and future prediction.
BibTeX
@inproceedings{lee2026physhandi,
  author = {Jihyun Lee and Changmin Lee and Donghwan Kim and Tae-Kyun Kim},
  title = {{PhysHanDI: Physics-Based Reconstruction of Hand-Deformable Object Interactions}},
  booktitle = {International Conference on Machine Learning},
  year = {2026},
  url = {https://openreview.net/forum?id=A93BskpqCP},
}
Figure