Research map / Graph learning / Knowledge and temporal graphs
EGODE: An Event-attended Graph ODE Framework for Modeling Rigid Dynamics
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 EGODE, a coupled graph ODE with a learnable event module for collision-triggered instantaneous updates, to model rigid-body dynamics at mesh-node and object levels.
Problem
Rigid-body dynamics simulation with GNN simulators struggles to combine continuous evolution with instantaneous collision changes and to model hierarchical mesh-node/object structure.
Approach
Couples a node-level and an object-level graph neural ODE and adds a learnable event detector that estimates collision times and applies discrete state updates at those times during ODE integration.
Key results
- On Physion, EGODE reaches 94.7% accuracy on Dominoes vs 88.8% for the best baseline SGNN, and 90.0% on Collide vs 85.5% for SEGNO.
- On RigidFall with 500 training samples, EGODE's MSE at t=20 is 0.17 (×10^-2) versus 0.32 for SGNN.
- Ablation removing the neural ODE (w/o O) causes the largest performance drop among the three components.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the neural ODE component (ablation) | EGODE · Physion contact prediction (eight scenarios) | contact prediction accuracy (%): 89.8±1.7 -> 94.7±1.4 (Dominoes); 86.3±1.2 -> 90.0±1.0 (Collide) | Helped |
| add the coupled architecture (ablation) | EGODE · Physion contact prediction (eight scenarios) | contact prediction accuracy (%): 90.7±1.0 -> 94.7±1.4 (Dominoes); 88.0±1.1 -> 90.0±1.0 (Collide) | Helped |
| add the event module for collision modeling (ablation) | EGODE · Physion contact prediction (eight scenarios) | contact prediction accuracy (%): 90.3±1.5 -> 94.7±1.4 (Dominoes); 86.9±0.7 -> 90.0±1.0 (Collide) | Helped |
| use EGODE instead of the strongest reported baseline | Physion · Physion contact prediction, all eight scenarios | contact prediction accuracy (%): 94.7±1.4 vs 88.8±2.0 (Dominoes); 90.0±1.0 vs 85.5±1.2 (Collide) | Helped |
| use EGODE instead of SGNN | RigidFall · RigidFall (|Train| = 5000, t = 20 and t = 40) | prediction MSE (×10−2): 0.19±0.19 -> 0.12±0.11 (t = 20); 0.74±1.37 -> 0.46±0.42 (t = 40) | Helped |
| set λ to 1 in the training objective | EGODE · Physion Dominoes and Collide | contact prediction accuracy | Helped |
| add resistive force proportional to velocity | EGODE · Physion Collision | motion under external force | Mixed |
Surprising findings
- Standard MSE is unreliable for some Physion scenes (Contain, Drop, Link, Support) because centrally symmetric motions are all plausible, so the authors rely on contact accuracy there.
Limitations
- Cannot handle rigid body hinges or deformable objects.
- Evaluated only on Physion and RigidFall benchmarks.
Experimental setup
Datasets: Physion RigidFall
Benchmarks: Physion RigidFall
Metrics: contact prediction accuracy mean square error MAE
Baselines: MLP RNN SocialODE GNS DPI-Net EGNN GMN SGNN SEGNO
Builds on: neural ODE SGNN GNS MeshGraphNet
Closest papers by meaning
- Learning Flexible Body Collision Dynamics with Hierarchical Contact Mesh Transformer (ICLR 2024 · similarity 0.85)
- Learning rigid-body simulators over implicit shapes for large-scale scenes and vision (NeurIPS 2024 · similarity 0.85)
- PGODE: Towards High-quality System Dynamics Modeling (ICML 2024 · similarity 0.85)
- Analytical Construction on Geometric Architectures: Transitioning from Static to Temporal Link Prediction (ICML 2025 · similarity 0.84)
- PhyMPGN: Physics-encoded Message Passing Graph Network for spatiotemporal PDE systems (ICLR 2025 · similarity 0.84)
- EvoMesh: Adaptive Physical Simulation with Hierarchical Graph Evolutions (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.