Contact-rich manipulation—tasks that require managing physical contact—is central to everyday robot use, but evaluating such skills is hard because contact behaviors vary widely and tasks are difficult to reproduce across labs. To address this, the authors of ContactBench 1.0 studied everyday human manipulation and distilled it into eight recurring skill categories. They then built an extensible benchmark that instantiates these skills through 36 reference tasks and more than 40 object assets.

The assets are designed to capture task-relevant geometry and contact behavior, and they can be fabricated with consumer-grade 3D printers. Each asset is paired with a simulation model that includes tunable resistance loads, and controlled task variations allow researchers to test how well policies generalize. This combination of physical and simulated resources is intended to make evaluation more consistent across different research groups.

The authors evaluated recent vision-based and vision-tactile policy learning methods on six of the benchmark tasks. The results show that the vision-only policies struggled to perceive or react to contact events, while the visual-tactile policies did not meaningfully improve performance. This suggests that current learning methods still have a significant gap when it comes to contact-rich manipulation.

By providing shared tasks, physical assets, and simulation models, ContactBench 1.0 aims to establish a foundation for reproducible evaluation in this area. The benchmark is open-sourced, allowing other researchers to use the same tasks and assets to compare future methods more directly.