Benchmarking Rigid
Body Contact Models

Michelle Guo, Yifeng Jiang, Andrew Spielberg, Jiajun Wu, Karen Liu

As robots are increasingly deployed in contact-rich tasks, there has been increased interest in models of contact that are more accurate than those of untuned simulations. These methods typically rely on simulators that have been system-identified, full dynamical models that are learned, or a combination of both approaches. These methods have typically targeted scenes with well-behaved physical parameters and a single body; however, wider ranges of phenomena are important for many real-world settings and serve as stress-tests that probe the strengths and weaknesses of these methods. In this study, we present a large synthesized dataset with diverse scenes, including objects with varying materials and geometries, or multiple objects involved in inter-body collisions. We use this dataset, to compare and contrast recent approaches in a systematic and unified way. Our empirical evaluations show that while some analytical methods work well in some settings and learned (and hybrid) methods work well in others, no existing method excels in all situations, and all tend to struggle as geometric complexity and the number of scene bodies increase. Our findings call for the collection of more diverse real-world contact datasets for better evaluation of future models.