METRICS/ADAPT Dataset: Sim2Real Object Detection and Pose Estimation

The ADvanced Agile ProducTion (ADAPT) competition, a part of the EU Horizon-2020 funded

project METRICS, is designed to address the challenges in dexterous manipulation of mechanical

parts within the assembly processes. Central to this competition are object detection and pose

estimation capabilities, which are pivotal in contemporary robotic manipulation systems and

which increasingly rely on machine learning algorithms. Hence, evaluating the performance of

these algorithms, particularly when their training is constrained by limited access to

real-world data, is crucial in assessing the readiness for deployment of such systems in

practical settings.

The ADAPT dataset was therefore created specifically for this purpose. It contains detailed

CAD models of three different assembly parts and a collection of real, annotated data. This

data is essential for testing how well part detectors and pose estimators perform in real

situations. Algorithms can be trained using synthetic images created from these CAD models,

thereby helping them learn in simulated conditions. This method is key in closing the gap

between training in simulations and working in actual environments, underscoring the dataset's

role in advancing the use of robots in industrial settings.

先进敏捷生产（ADAPT）竞赛作为欧盟地平线2020（EU Horizon-2020）资助项目METRICS的组成部分，旨在解决装配流程中机械部件灵巧操作的相关挑战。本次竞赛的核心关注点为目标检测与位姿估计能力——此类能力在当代机器人操作系统中至关重要，且愈发依赖机器学习算法。因此，评估此类算法的性能，尤其是在训练受限于有限真实世界数据获取的场景下，对于判断这类系统能否在实际场景中部署至关重要。
为此，ADAPT数据集专为该竞赛需求定制。数据集包含三种不同装配部件的详细计算机辅助设计（CAD）模型，以及一批带标注的真实采集数据。此类数据可用于测试部件检测器与位姿估计器在真实场景中的实际表现。研究人员可借助这些CAD模型生成的合成图像对算法开展训练，使其在仿真环境中完成学习。该方法是缩小仿真训练与实际部署性能差距的核心路径，也凸显了本数据集在推动工业场景机器人应用发展中的关键作用。