With advancements in deep learning, robots have significant potential as embodiments of artificial intelligence in both factory automation and daily life applications. To effectively interact with the real world, accurate environmental perception in 3D space is crucial for subsequent manipulation tasks. Specifically, determining the rotation and translation of objects, known as the 6D object pose estimation, is a critical step in enabling precise robotic interactions.
To address 6D object pose estimation, traditional methods such as SIFT and point pair features are commonly used to track key features and estimate object poses. However, these approaches exhibit limited performance in real-world applications. Recent deep learning-based methods have significantly improved pose estimation accuracy, but they typically support only a limited set of object instances. In addition to this generalization limitation, deep learning approaches require extensive training datasets with 6D pose annotations, which are costly and time-consuming to collect. Consequently, a gap remains between current research and the development of a generalizable pose estimator which can be used off-the-shelf.
Firstly, we aim to reduce the complexity of object CAD scanning and model training. To this end, we propose an in-hand object onboarding pipeline, followed by a novel object pose estimator, which simplifies the training process and reduces its requirements.
Our second focus is on transparent and metallic objects, which are common in household environments but often neglected in current pose estimation methods. To address this, we introduce a polarization modality and fuse multi-modal information to overcome depth measurement challenges on photometrically-challenging object surfaces. Thirdly, we focus on reducing the large amount of real data needed in the pose training. It is a natural idea to leverage synthetic images to meet the training data requirements, but the domain gap between synthetic and real images significantly impacts network performance. To mitigate this, our primary focus is on Sim2Real domain adaptation, which relies solely on synthetic images during training. In particular, we incorporate 2D foundation model features into the object point cloud and fuse both semantic and geometric features in the network, significantly enhancing generalization to real-world images.
«
With advancements in deep learning, robots have significant potential as embodiments of artificial intelligence in both factory automation and daily life applications. To effectively interact with the real world, accurate environmental perception in 3D space is crucial for subsequent manipulation tasks. Specifically, determining the rotation and translation of objects, known as the 6D object pose estimation, is a critical step in enabling precise robotic interactions.
To address 6D object pose...
»