Research map / Video / Video generation
POMP: Physics-consistent Motion Generative Model through Phase Manifolds
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes POMP, a real-time autoregressive motion generator that pairs a phase-guided diffusion kinematic module with a PhysX-in-Unity dynamic module, using an aligned periodic autoencoder in a phase manifold to bridge the kinematic-simulated domain gap.
Problem
Kinematic real-time motion generators produce physically implausible artifacts (foot sliding, floating, penetration), and naively adding a physics simulator to a pretrained kinematic prior causes the simulated poses to drift off the motion prior, leading to generation failure.
Approach
Insert a phase encoding module (aligned-PAE) that projects simulated poses back onto the shared phase manifold of the kinematic motion prior, matching motion type while re-timing within the same phase ellipse, so a non-differentiable physics simulator can sit outside the training loop.
Key results
- On the 100STYLE setting, POMP reports FMD 1.02 and foot skating 0.06, versus 2.31 for StyleERD and 3.15 for RTSM (ranked best FMD among 100-style methods except human-terrain variant).
- Removing the dynamic module raises foot penetration from 0% (POMP) to 12.7% on complex terrains, compared with 14.4% for PFNN.
- On AMASS, POMP reaches 100% test success with MPJPE 29.8 mm, versus PHC 99.2% / 36.1 mm and PULSE 97.1% / 54.1 mm; training takes 17.2 hours on a single RTX 3090.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use POMP instead of PFNN | Human-terrain motion generation · human-terrain | FMD: 10.25 -> 0.92 | Helped |
| Use POMP instead of RTSM | 100STYLE motion generation · 100STYLE | FMD: 3.15 -> 1.02 | Helped |
| Use POMP instead of PHC | AMASS · AMASS | test MPJPE: 36.1 -> 29.8 | Helped |
| Add the dynamic module (ablation) | POMP · complex terrains | Penetration: 12.7% -> 0% | Helped |
| Add PEM phase-manifold alignment (ablation) | POMP · walking-to-traversal transition | velocity trajectory alignment | Helped |
| Add Ortho-MoE instead of standard MoE (ablation) | POMP · human-terrain | FMD: 1.42 -> 0.92 | Helped |
| Use aligned-PAE to align simulated and kinematic poses (ablation) | POMP · walking-to-obstacle-crossing transition | phase representation | Helped |
Surprising findings
- Without the dynamic module, POMP's penetration rate (12.7%) offers little advantage over PFNN (14.4%).
Limitations
- Physics simulator is kept as a fixed non-differentiable component outside the training loop, so the kinematic module does not learn from simulation gradients.
- Larger compute and parameter count than PFNN (55.46M vs 2.39M params for the 100STYLE variant).
Experimental setup
Datasets: terrain-fitting dataset (Holden et al.) 100STYLE AMASS
Benchmarks: terrain-fitting human-terrain dataset 100STYLE AMASS
Metrics: FMD foot skating runtime #params penetration success rate MPJPE
Baselines: PFNN Mason et al. StyleERD RTSM Motion Puzzle PHC MaskedMimic PULSE
Builds on: PFNN PAE DROP Analytic-DPM DDIM
Closest papers by meaning
- KinemaDiff: Towards Diffusion for Coherent and Physically Plausible Human Motion Prediction (ICLR 2026 · similarity 0.87)
- PhyMotion: Structured 3D Motion Reward for Physics-Grounded Human Video Generation (NeurIPS 2026 · similarity 0.86)
- PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation (NeurIPS 2025 · similarity 0.85)
- Optimal-state Dynamics Estimation for Physics-based Human Motion Capture from Videos (NeurIPS 2024 · similarity 0.85)
- MotionMap: Representing Multimodality in Human Pose Forecasting (CVPR 2025 · similarity 0.85)
- Scaling Large Motion Models with Million-Level Human Motions (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.