REASSEMBLE: A Multimodal Dataset for Contact-rich Robotic Assembly and Disassembly

Motion Policy Learning 1. Idle Temporal Action 2. Pick Ethernet In REASSEMBLE, we focus on creating a dataset for contact-rich manipulation tasks. We leverage the well-established NIST Assembly Task Board #1 [1] to facilitate deployment of learned algorithms across different research institutes. The dataset includes data from various modalities, such as RGB cameras and robot proprioception, which are common in other works. Additionally, we incorporate event cameras, a force and torque sensor, and microphones, which are less common in manipulation datasets, and which we hope will be beneficial for the community. Abstract--Robotic manipulation remains a core challenge in advancing robotic manipulation in complex, real-world scenarios. In contrast, more complex challenges, such as longhorizon assemblies, and nut threading, as shown in Figure 1. These and contact-rich manipulation tasks, remain less explored. What sets REASSEMBLE which require understanding of interaction dynamics and apart from other robot manipulation datasets is its focus on the ability to plan and execute precise, goal-oriented action multimodal data for holistic learning frameworks. Recent studies, including [11], have highlighted a comparison of commonly used robot learning datasets the limitations of current state-of-the-art algorithms to address and their properties in Table I.