Research map / Probabilistic and scientific modeling / Molecules and proteins
Human Motion Prediction Under Unexpected Perturbation
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 Latent Differentiable Physics (LDP), combining an explicit differentiable inverted pendulum model with CVAE-based skeleton restoration, to predict human motion under unexpected physical perturbation in single- and multi-person settings from very little data.
Problem
Predicting human reactive motions under unexpected physical perturbation, including how pushes propagate through groups, where data is extremely scarce and interactions/forces are unobserved.
Approach
Maps full-body poses into a low-dimensional latent inverted pendulum model (IPM) state, simulates it with a differentiable physics model (balance-recovery PD control, friction, learned interaction forces), and recovers full-body poses with CVAE-based restoration networks, giving data efficiency and explainability.
Key results
- On single-person prediction, LDP improves MPJPE by up to 25.38% over the best baseline (MPJPE 0.097 vs siMLPe 0.130).
- On multi-people prediction, LDP outperforms all baselines by at least 34.57% on MPJPE (0.106 vs best 0.162 for MRT).
- Ablation shows removing the IPM guidance degrades MPJPE from 0.110 to 0.217 in the full-body setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add differentiable IPM guidance instead of predicting poses directly (ablation) | LDP · single-person | MPJPE: 0.217 -> 0.110 | Helped |
| Add differentiable IPM guidance instead of predicting poses directly (ablation) | LDP · single-person | FSE: 0.196 -> 0.126 | Helped |
| Separate lower- and upper-body pose reconstruction (ablation) | LDP with IPM · single-person | FSE: 0.126 -> 0.069 | Helped |
| Separate lower- and upper-body pose reconstruction (ablation) | LDP without IPM · single-person | FSE: 0.196 -> 0.313 | Hurt |
| Use LDP rather than the strongest baseline | single-person prediction · single-person | MPJPE: 0.130 -> 0.097 | Helped |
| Use LDP rather than the strongest baseline | multi-people prediction · multi-people | MPJPE: 0.162 -> 0.106 | Helped |
Surprising findings
- Removing the IPM guidance makes separate lower/upper body restoration ineffective, with FSE getting worse (0.313 vs 0.069 with IPM).
Limitations
- Requires explicit physical modelling, making it less general than black-box networks that are plug-and-play on data.
- Interaction forces between people are not measured in the data and must be learned as latent variables.
Experimental setup
Datasets: FZJ Push
Benchmarks: FZJ Push
Metrics: MPJPE hipADE hipFDE MBLE FSE
Baselines: A2M ACTOR MDM RMDiffuse PhyVae siMLPe EqMotion MRT DuMMF TBIFormer JRFormer
Builds on: inverted pendulum model differentiable physics CVAE siMLPe EqMotion
Closest papers by meaning
- Physical Plausibility-aware Trajectory Prediction via Locomotion Embodiment (CVPR 2025 · similarity 0.86)
- Existence Is Chaos: Enhancing 3D Human Motion Prediction with Uncertainty Consideration (AAAI 2024 · similarity 0.85)
- LAL: Enhancing 3D Human Motion Prediction with Latency-aware Auxiliary Learning (CVPR 2025 · similarity 0.85)
- PhysPT: Physics-aware Pretrained Transformer for Estimating Human Dynamics from Monocular Videos (CVPR 2024 · similarity 0.84)
- SimpliHuMoN: Simplifying Human Motion Prediction (NeurIPS 2026 · similarity 0.84)
- Homogeneous Dynamics Space for Heterogeneous Humans (CVPR 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.