Hybrid Architectures for Object Pose and Velocity Tracking at the Intersection of Kalman Filtering and Machine Learning

The study of object perception algorithms is fundamental for the development of robotic platforms capable of planning and executing actions involving objects with high precision, reliability and safety. Indeed, this topic has been vastly explored in both the robotic and computer vision research communities using diverse techniques, ranging from classical Bayesian filtering to more modern Machine Learning techniques, and complementary sensing modalities such as vision and touch. Recently, the ever-growing availability of tools for synthetic data generation has substantially increased the adoption of Deep Learning for both 2D tasks, as object detection and segmentation, and 6D tasks, such as object pose estimation and tracking. The proposed methods exhibit interesting performance on computer vision benchmarks and robotic tasks, e.g. using object pose estimation for grasp planning purposes. Nonetheless, they generally do not consider useful information connected with the physics of the object motion and the peculiarities and requirements of robotic systems. Examples are the necessity to provide well-behaved output signals for robot motion control, the possibility to integrate modelling priors on the motion of the object and algorithmic priors. These help exploit the temporal correlation of the object poses, handle the pose uncertainties and mitigate the effect of outliers. Most of these concepts are considered in classical approaches, e.g. from the Bayesian and Kalman filtering literature, which however are not as powerful as Deep Learning in handling visual data. As a consequence, the development of hybrid architectures that combine the best features from both worlds is particularly appealing in a robotic setting. Motivated by these considerations, in this Thesis, I aimed at devising hybrid architectures for object perception, focusing on the task of object pose and velocity tracking. The proposed architectures use Kalman filtering supported by state-of-the-art Deep Neural Networks to track the 6D pose and velocity of objects from images. The devised solutions exhibit state-of-the-art performance, increased modularity and do not require training to implement the actual tracking behaviors. Furthermore, they can track even fast object motions despite the possible non-negligible inference times of the adopted neural networks. Also, by relying on data-driven Kalman filtering, I explored a paradigm that enables to track the state of systems that cannot be easily modeled analytically. Specifically, I used this approach to learn the measurement model of soft 3D tactile sensors and address the problem of tracking the sliding motion of hand-held objects.