Keskin, Esila, Yang, Xintong  ORCID: https://orcid.org/0000-0002-7612-614X, Wei, Minglun and Ji, Ze
2026.
Event-based physics parameter inference from GPU-accelerated synthetic simulations using convolutional neural networks.
Presented at: 2026 31st International Conference on Automation and Computing (ICAC),
High Wycombe, UK,
26 - 28 August 2026.
2026 31st International Conference on Automation and Computing (ICAC).
IEEE,
10.1109/icac70325.2026.11712415
|
Preview |
PDF
- Accepted Post-Print Version
Available under License Creative Commons Attribution. Download (849kB) | Preview |
Abstract
Identifying physical properties of objects from visual observations is a fundamental challenge in robotics. This paper presents a GPU-accelerated synthetic physics simulation pipeline using NVIDIA Warp, generating large-scale labelled datasets with domain-randomised physics parameters including gravity, restitution, and Coulomb friction. We investigate four neural network architectures for physics parameter inference from visual input: a single-frame Convolutional Neural Network (CNN), a two-frame CNN, a single event-frame CNN inspired by neuromorphic sensing, and a 5-frame sequence event CNN. Our primary contribution is demonstrating that environment-level physics parameters can be inferred directly from visual observations using a temporal event-based representation. Event-based encoding of frame differences outperforms raw frame input while using significantly fewer input channels, and temporal context across multiple frames further improves inference. We evaluate noise robustness, RGB+Event fusion, dedicated single-parameter inference heads, recurrent and structural temporal baselines, differentiable simulation, and an extended rolling friction setting with angular velocity prediction. Results confirm that sparse temporal motion signals are more informative than static or short-range RGB input for physics parameter inference across three datasets of increasing physical complexity.
| Item Type: | Conference or Workshop Item - published (Paper) |
|---|---|
| Date Type: | Published Online |
| Status: | Published |
| Schools: | Schools > Engineering |
| Additional Information: | RRS policy applied |
| Publisher: | IEEE |
| Date of First Compliant Deposit: | 13 October 2026 |
| Last Modified: | 13 Oct 2026 13:45 |
| URI: | https://orca.cardiff.ac.uk/id/eprint/190144 |
Actions (repository staff only)
 |
Edit Item |
Dimensions
Dimensions