Recent breakthroughs in generative AI have enabled the synthesis of complex 3D human movements from simple text prompts or music. However, a persistent "uncanny valley" remains: while these generated motions often look smooth, they frequently violate the laws of physics. Characters "skate" across the floor, limbs clip through solid objects, and bodies appear to float without weight.
The research community is currently shifting from purely kinematic models—which focus on the geometry of joint angles—to physically grounded frameworks that account for forces, contacts, and environmental constraints.
The Kinematic Gap
Most state-of-the-art motion models, such as TEMOS by Mathis Petrovich, Gül Varol, and Michael J. Black, utilize Variational Autoencoders (VAEs) or Diffusion models to map high-level descriptions to sequences of joint rotations. While successful at capturing the style of a motion, these models are often "physics-blind."
Because the training objective typically prioritizes minimizing the distance between predicted and ground-truth joint positions, the model may find a solution that is visually similar but physically impossible. For example, in Transflower, Guillermo Valle-Pérez and Simon Alexanderson note that without explicit physics constraints, even high-quality dance generation suffers from "foot sliding," where the feet move slightly while supposedly planted on the ground. This happens because the model has no concept of friction or ground reaction forces.
Grounding Motion in Contact
To solve these artifacts, researchers are integrating contact modeling directly into the generative pipeline. The intuition is that motion is not just a series of poses, but a series of interactions with the environment.
- Stochastic Interaction: In SAGA, Jiahao Wang and Otmar Hilliges address whole-body grasping by first generating a static "contact-rich" end-pose and then infilling the motion. By anchoring the motion to a physical contact point (the object), the resulting movement is more plausible.
- Sparse Tracking: The AGRoL framework by Yuming Du and Sebastian Starke uses diffusion to predict full-body motion from just head and hand trackers. They emphasize that integrating physical priors is necessary to prevent the "floating" effect common in VR avatars.
- Implicit Representations: Models like TOHO, developed by Quanzhou Li and Bo Dai, use Implicit Neural Representations (INRs) to treat motion as a continuous function. While this allows for smooth upsampling, the authors acknowledge that without an integrated physics solver, these models still struggle with hand-object penetration.
From Frames to Functions and Forces
A newer direction involves rethinking how we represent motion entirely. Instead of a sequence of discrete frames, researchers like Marco Pegoraro and Evan Atherton propose FunPhase, which represents motion as a continuous mathematical function within a "phase manifold."
While these functional representations improve visual smoothness, the field is converging on the idea that visual fidelity is not enough. As Han Liang and colleagues argue in their work on OMG, purely kinematic models must eventually be paired with physics-based simulators or reinforcement learning to ensure that movements adhere to gravity and momentum.
Iterative Refinement and Reasoning
The most recent trend involves treating physical plausibility as a "reasoning" problem. IRG-MotionLLM, proposed by Qize Yang and Ling-An Zeng, introduces an iterative loop where the model generates a motion, assesses its own physical and semantic quality, and then refines it.
Similarly, AvatarGO by Yukang Cao and Kwan-Yee K. Wong uses Large Language Models (LLMs) to reason about the spatial relationships in human-object interactions. By understanding that a "heavy" object requires different balance than a "light" one, these models begin to bridge the gap between linguistic intent and physical reality.
The ultimate goal, as summarized in the survey by Aliasghar Khani and Arianna Rampini, is a unified framework where neural generation provides the creativity and variety of human movement, while differentiable physics engines provide the "hard constraints" of the real world.