Abstract :
[en] Robotic grasping is the task of picking up an object by applying forces and torques at a set of contacts with a gripper. This well-known problem, with more than four decades of research, still poses significant challenges mainly for three reasons: the rich nonsmooth contact dynamics of grasping, the high dimensional search space of grippers, and the sensor noise. Physical modeling provides a mathematical description of the contact dynamics, leading to the first approach in robotic grasping. Then, the profusion of data due to the rapid progress of information technology has made machine learning, a sub-field of artificial intelligence, an uncontested approach for solving real-world problems. It was quickly applied to robotic grasping with incredible results. However, both approaches neglect one of the most fundamental issues: dealing with uncertainties. In this thesis, we aim to develop a general probabilistic framework, describing the grasping problem with random variables, to infer the grasp pose thanks to the Bayes' rule, which is a principled approach for dealing with uncertainties: our prior beliefs are updated based on new available observations. To this end, we structure our work with four contributions. The first contribution is to lay down the foundation of our approach. We develop and explain the key components. The first component consists of the probabilistic modeling of the variables used in robotic grasping. We describe the quantities needed to solve the task by incorporating human-based knowledge of grasping while being general enough to make our modeling applicable in many situations. The second component is the method used to carry out the inference. Due to the intractability of the likelihood function, we use a family of methods known as simulation-based inference. These methods learn a part of the Bayes' rule using neural networks as surrogate models. However, they typically require lots of data, which is very costly to gather in robotics. Therefore, we leverage robotic simulators to generate samples representing the task of grasping. The last component consists of using geometric methods to search for the grasp pose by performing a Riemannian gradient descent scheme, which preserves the geometry of the space of some variables. The next three contributions consist of designing priors relevant to the constraints of the task. While a very simple prior is used in the first contribution, bringing almost no information about the task, we propose to use invariance as a good property to inject into the prior in the second contribution. Thus, it becomes possible to complexify the task by adding more unknown parameters. These priors, however, are based on the object. To overcome this issue, we use implicit neural representations to model a prior based on the scene in the third contribution, which allows one to reason about the environment and not only one object. We extend these capabilities by providing a systematic way of designing priors in the last contribution, thus achieving a high success rate on complex tasks.
