Abstract

While current general-purpose 3D human models (e.g., SMPL-X) efficiently represent accurate human shape and pose, they lack the ability to physically interact with the environment due to their kinematic nature. This results in kinematic-based interaction models often suffering from issues such as interpenetration and unrealistic object dynamics.

We propose Half-Physics, a mechanism that embeds SMPL-X into a tangible entity capable of dynamic physical interactions with its surroundings. Instead of applying torques (as in RL-based methods), we explicitly enforce velocities computed from kinematic targets, maintaining kinematic control over inherent SMPL-X poses while ensuring physically plausible interactions. Unlike reinforcement learning-based methods, Half-Physics is learning-free, generalizes to any body shape and motion, and operates in real time (952 fps).

Method Overview

The half-physics approach transforms discrete kinematic poses into continuous velocity representations:

1. Physics-compatible SMPL-X: Create a variant with 55 rigid body parts suitable for physics simulation.
2. Velocity computation: Compute target velocities from kinematic frames using finite differences.
3. Physics simulation: Execute simulation with explicit velocity constraints using the Bullet physics engine.
4. Passive Joint Stiffness Compensation (PJSC): Dynamically apply torques during substeps to handle collisions while preserving kinematic intent.
5. Inverse kinematics: Update joint states post-simulation for consistency.

Comparison with RL-based Methods

Unlike PHC+ which can fail or accumulate tracking errors, Half-Physics maintains perfect fidelity to the kinematic reference in collision-free scenarios, with graceful degradation only where physics constraints require it.

Human-Scene Interaction

On the Trumans dataset, Half-Physics eliminates all penetration artifacts:

• Penetration rate: 7.91% → 0%
• Average penetration depth: 82.10 mm → 0 mm
• Maximum penetration: 172.19 mm → 0 mm

Human-Object Interaction

Half-Physics enables diverse physically plausible interactions including single-handed grasping, two-handed manipulation, object dropping with realistic gravity, and kicking with motion-magnitude-aware responses. It also enables data augmentation by varying physical properties (mass, friction) to generate diverse interaction outcomes.

Key Advantages

• Training-free: No learning required; works immediately with any motion
• Real-time: 952.94 fps on standard desktop hardware
• High fidelity: Preserves original kinematic motion in collision-free scenarios
• Generalization: Applies uniformly across body shapes and motion types
• Physically correct: Objects do not move before actual contact, unlike purely kinematic systems

BibTeX

@article{siyao2025halfphysics,
  title   = {Half-Physics: Enabling Kinematic {3D} Human Model with Physical Interactions},
  author  = {Siyao, Li and Feng, Yao and Taheri, Omid and Loy, Chen Change and Black, Michael J.},
  journal = {arXiv preprint arXiv:2507.23778},
  year    = {2025},
}