Physically Plausible Generative Human Motion

This direction is merged into another direction: foot sliding artifacts. Its last version stays readable and citable here.

Version 1 of 1284 papers314 conceptsfrozen 2026-10-04

Generative models for human motion have reached a point where they can synthesize complex dances, athletic maneuvers, and intricate interactions from simple text prompts. However, a persistent "uncanny valley" remains: the generated characters often appear to glide across the floor, hover in mid-air, or pass their limbs through solid objects. This discrepancy arises because most models are purely kinematic—they learn the statistical patterns of joint positions over time without understanding the underlying forces, like gravity and friction, that make movement possible.

Recent research has shifted from simply generating "visual-looking" motion to enforcing physical grounding, ensuring that every step and jump adheres to the laws of physics.

The Kinematic Gap and the Foot Sliding Problem

The most common artifact in neural motion synthesis is "foot sliding" (or foot skating). In a real walk, the foot in contact with the ground has zero velocity relative to the floor. However, as noted in Human Motion Generation: A Survey, generative models like Variational Autoencoders (VAEs) and Diffusion Models often fail to maintain these strict contact constraints.

Xin Chen and colleagues demonstrate in Executing your Commands via Motion Diffusion in Latent Space that even when models operate in efficient latent spaces to improve smoothness, they do not inherently guarantee adherence to physical laws. The model might predict a visually graceful arc for a leg, but without a concept of friction, the supporting foot "skates" to compensate for the momentum, breaking the illusion of weight.

Closing the Loop with Physics Simulation

To bridge this gap, researchers are increasingly "closing the loop" between flexible generative models and rigid physics engines. Instead of letting a diffusion model have the final say, its output is treated as a "plan" that a physics-based controller must then execute.

In CLoSD: Closing the Loop between Simulation and Diffusion, Guy Tevet and Zhengyi Luo use a "Diffusion Planner" to suggest motions which are then filtered through a Reinforcement Learning (RL) controller. This ensures the character can only move by applying simulated joint torques, making artifacts like floor-penetration or floating physically impossible. Similarly, QuestSim by Alexander Winkler and Yuting Ye uses this torque-driven approach to track full-body motion from sparse VR sensors, effectively "filling in" missing leg data by simulating how a physical body would have to balance itself.

Learning Physical Commonsense via Reinforcement

Running a full physics simulator during inference is computationally expensive and often non-differentiable, making it hard to train models directly. A more recent trend involves fine-tuning generative models using "physical commonsense" rewards.

Gaoge Han and the team behind ReinDiffuse treat the diffusion process as an action distribution. By applying RL rewards that penalize foot sliding and floating, the model learns to internalize physical constraints during the denoising process itself. This allows for high-fidelity motion that respects the floor without the need for a heavy simulator at runtime.

Embedding Biomechanics into the Architecture

Beyond external simulators or rewards, some researchers are embedding physical equations directly into the neural architecture. Zixi Kang introduces BioMoDiffuse, which incorporates Euler-Lagrange equations into the generative process. By treating the human body as a dynamic system driven by muscle activations and forces, the model moves away from simple coordinate regression toward a representation that is biologically and mechanically authentic.

Why Scaling Data Isn't Enough

A natural question is whether massive datasets can simply "teach" physics to a model through observation. Evidence from Go to Zero by Ke Fan suggests otherwise. Even when scaling to 7-billion parameters and 2 million motion sequences, artifacts like gravity defiance persist. This indicates that physical plausibility is not just a matter of model capacity, but a structural requirement. As long as training data contains noise or lacks explicit contact labels, models will continue to produce "floaty" results unless they are explicitly grounded by physical priors or simulation-based constraints.

Papers in this version

Showing 60 of 284 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_physically_plausible_generative_human_mo_2010613e,
  title = {Physically Plausible Generative Human Motion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physically-plausible-generative-human-motion/rdr_2010613e170840e19c2a1c8b63d1a9fd},
  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_physically_plausible_generative_human_mo_2010613e_v1,
  title = {Physically Plausible Generative Human Motion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/physically-plausible-generative-human-motion/rdr_2010613e170840e19c2a1c8b63d1a9fd/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}