Physics Constrained Generative Human Motion Synthesis

2,028 concepts1,054 papers20 proposals

Modern deep generative models excel at translating abstract inputs—like natural language or style references—into complex 3D human animations. Frameworks like the Transformer-based VAE proposed by Mathis Petrovich et al. in TEMOS: Generating Diverse Human Motions from Textual Descriptions capture the rich semantic diversity of human movement. However, a persistent flaw plagues these purely data-driven approaches: they frequently produce physical anomalies such as foot sliding (or "foot skating"), ground penetration, and floating.

The Kinematic-Dynamic Gap

Why do highly sophisticated generative models produce such basic physical errors? The root cause lies in the distinction between kinematics and dynamics.

Traditional generative architectures operate strictly in the kinematic domain. They learn to predict joint positions, velocities, or angles by mimicking statistical distributions found in motion capture datasets. As noted by Wanying Zhang et al. in Dynamic Compositional Graph Convolutional Network for Efficient Composite Human Motion Prediction, these models prioritize temporal smoothness over physical validity.

Because a neural network has no intrinsic concept of gravity, mass, or friction, it treats a foot planted firmly on the ground the same way it treats a hand waving in the air. When a model blends different styles or interpolates between poses, the strict mathematical constraints of a static contact point are lost. This results in characters that appear to glide frictionlessly across the floor, a common issue highlighted by Lei Zhong in SMooDi: Stylized Motion Diffusion Model and Aman Goel et al. in Interaction Mix and Match: Synthesizing Close Interaction using Conditional Hierarchical GAN with Multi-Hot Class Embedding.

To bridge this kinematic-dynamic gap, researchers have developed mechanisms that enforce physical laws during or after the generation process.

Anchoring Motion with Contact Constraints

To eliminate artifacts like foot sliding, models must understand exactly when a limb should be stationary. If the network is aware of contact states, it can mathematically anchor the joints to the environment.

An intuitive way to implement this is through binary environmental interaction labels. Shihong Xia et al. demonstrated in Combining Recurrent Neural Networks and Adversarial Training for Human Motion Synthesis and Control that explicitly modeling contact states—predicting whether a foot is touching the ground at any given frame—forces the generator to respect the discrete transitions of locomotion. When the contact label is active, the loss function heavily penalizes any velocity at the foot joint, effectively gluing it to the floor and preventing skating.

[Generative Network] ---> Predicts Kinematic Poses + Contact Labels (0 or 1)
                                                       |
                                                       v
                                        [Zero-Velocity Loss Function]
                                       (Penalizes joint movement when contact = 1)

This concept extends naturally to hand-object and character-scene interactions. For instance, Junuk Cha et al. in Text2HOI: Text-Guided 3D Motion Generation for Hand-Object Interaction utilize a static contact map as a geometric prior before generating dynamic motion. By pre-determining exactly where a hand will touch an object, the downstream diffusion model is constrained to generate trajectories that land precisely on those points, preventing hands from drifting or clipping through solid surfaces (Haziq Razali et al. in Action-Conditioned Generation of Bimanual Object Manipulation Sequences).

Physics Engines and RL as Differentiable Filters

While contact maps prevent local drifting, they do not guarantee global physical plausibility. A character might keep their feet glued to the floor but still lean at an angle that violates balance and gravity. To resolve this, researchers treat physics as a non-negotiable filter or regularizer.

Instead of forcing a neural network to learn the laws of physics from scratch, frameworks like LAGOON, developed by Shusheng Xu et al., decouple the pipeline into a two-stage process:

  1. Kinematic Proposal: A physics-agnostic generative model proposes a visually realistic reference animation based on text or style.
  2. Dynamic Projection: A reinforcement learning (RL) agent operating inside a high-fidelity physics engine attempts to track that reference trajectory.

Because the RL agent must obey simulated gravity and friction to avoid falling, it naturally filters out physically impossible movements. The agent finds the closest viable trajectory on the physical manifold that matches the visual intent of the reference animation.

[Text Prompt] -> [Diffusion Model] -> [Impossible Kinematic Reference]
                                                    |
                                                    v
                                        [Physics Engine / RL Agent]
                                                    |
                                                    v
                                        [Physically Viable Trajectory]

Similarly, the authors of PoseAugment utilize a physical optimization module that forces noisy kinematic outputs to adhere to equations of ground reaction forces and torques. This optimization step strips away high-frequency jitter and foot sliding, transforming raw statistical generations into physically grounded sequences.

Conditioning on Intrinsic Physical Properties

A deeper reason kinematic models fail is that they treat all objects and environments as hollow geometric meshes. In reality, how a human moves depends entirely on the physical properties of the world around them—such as mass, friction, and torque.

To address this, Soshi Shimada et al. introduced MACS: Mass Conditioned 3D Hand and Object Motion Synthesis, which explicitly conditions a generative diffusion model on the mass of an object. When the model knows an object is heavy, the synthesized hand trajectories naturally adapt, switching from delicate fingertip grips to full-palm grasps. Without this explicit conditioning, movements appear floaty and disconnected from reality, as noted by Chongkai Gao et al. in FLIP: Flow-Centric Generative Planning for General-Purpose Manipulation Tasks.

By transitioning from purely kinematic pattern matching to physics-aware architectures, the field is closing the gap between visual imagination and physical reality, ensuring that generated characters move with the weight, balance, and groundedness of the real world.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_foot_sliding_artifacts_d5f06bb7,
  title = {foot sliding artifacts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/foot-sliding-artifacts/rdr_d5f06bb72f8741fea69d836ae254f7fb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_foot_sliding_artifacts_d5f06bb7_v1,
  title = {foot sliding artifacts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/foot-sliding-artifacts/rdr_d5f06bb72f8741fea69d836ae254f7fb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake