In recent years, robot learning has made significant progress, with advances in general manipulation capabilities. Solutions using foundation models and vast collections of human demonstrations have shown strong generalization. Additionally, pretrained vision/language models have leveraged their common sense knowledge to perform various manipulation tasks. However, despite these achievements, the dexterity of these systems remains limited, typically only able to perform simple pick-and-place actions, falling short of the contact-rich skills required for more complex industrial and household tasks, such as insertion, assembly, cooking, or using tools.
On the other hand, recent work in model-based control and planning has achieved impressive dexterity in contact-rich tasks. However, these approaches struggle with the generalization and scalability that data-driven methods offer. This workshop aims to bring together experts from both academia and industry to foster discussions about how to integrate these two paradigms, ultimately enabling robots to perform a wider range of contact-rich tasks in real-world settings. The goal is to bridge the gap and ensure that robots can operate more robustly and adaptively in the open world.
