Learning Data-Efficient Rigid-Body Contact Models: Case Study of Planar Impact

Nima Fazeli, Samuel Zapolsky, Evan M. Drumwright, Alberto RodriguezCoRL 2017

Introduction

Rigid-body contact models are the mathematical backbone of modern physics engines used in robotics for motion planning and controller design. However, traditional analytical models often fail to capture the complexities of real-world impacts due to simplifying assumptions like point contact and instantaneous impact. This research explores whether data-driven approaches, constrained by physical principles, can bridge the gap between simulation and reality.

The Energy Ellipse and Physics-Inspired Learning

To improve predictive accuracy without requiring massive datasets, the authors leverage the "Energy Ellipse" concept. This geometric representation defines the space of all impulses that satisfy fundamental physical constraints, such as conservation of energy and non-penetration.

The Energy Ellipse provides a geometric constraint for admissible impulses, serving as a foundation for feature engineering in learned models.

By using the Energy Ellipse for feature engineering, the input state is reduced from a high-dimensional vector to a more compact representation. This allows Gaussian Process (GP) models to learn the mapping from pre-impact states to post-impact velocities with high data efficiency. The study compares standard analytical models against these learned models, categorized as either purely data-driven (Class I) or data-reinforced (Class II), which use physics as a strong prior.

Experimental Validation

The researchers developed a custom planar impact rig to collect high-fidelity ground truth data. An elliptical object was dropped between glass sheets, and its motion was captured at 250Hz. Over 1,700 valid trials were recorded, providing a robust empirical dataset for comparing analytical and learned models. This dataset is significant because it provides empirical evidence in a field that often relies on purely theoretical benchmarks.

Key Results and Convergence

The study found that learned models significantly outperform even the best-tuned analytical models. A key insight is that the structural limitations of classical equations—not just parameter tuning—prevent them from capturing phenomena like back-spin and non-centralized forces.

The error surface (a) and parameter convergence (b) demonstrate that learned models achieve high accuracy and stable parameter estimation with as few as 100 contact events.

As shown in the results, the models converge rapidly, reaching optimal performance with roughly 100 samples. This efficiency makes them practical for real-world robotic applications where data collection can be expensive. Both the data-driven and data-reinforced models outperformed the "IRB Bound," the theoretical limit for any model following standard rigid-body assumptions.

Conclusion

By combining the expressivity of Gaussian Processes with the constraints of rigid-body dynamics, this work demonstrates a path toward more accurate and data-efficient contact modeling. While current models were tested on a single elliptical geometry, future research will focus on generalization to diverse object shapes and the integration of these models into real-time robotic controllers.

Summary updated

BibTeX
@inproceedings{fazeli2017learning,
  author = {Nima Fazeli and Samuel Zapolsky and Evan M. Drumwright and Alberto Rodriguez},
  title = {{Learning Data-Efficient Rigid-Body Contact Models: Case Study of Planar Impact}},
  booktitle = {Conference on Robot Learning},
  year = {2017},
}
Figure