Abstract
Robotic grasp of cluttered scene has remained one of the most basic issues, primarily due to the widespread obstructions, sensor noise and natural deficiency in full view information. Modern vision-based manipulation systems heavily rely on more confident perception pipelines far more often than indeterminate perception pipelines are capable of addressing the uncertainty of object pose recognition and scene understanding, thus leading to fragile grasp execution under operational conditions. In this paper I will introduce an uncertainty conscious and vision based robotic manipulation system explicitly modelling and exploiting perceptual uncertainty when planning grasps in cluttered environments. This method combines a probabilistic visual perception and learning based grasp generation which enables the robot to reason in the face of uncertainty embedded in occlusions, depth ambiguity and finite expressiveness of the used models. We build indeterministic representations of the objects and grasps by applying uncertainty-estimation methods in the visual pipeline and utilize these representations of the objects and grasps in a grasp-selection strategy that is risk-sensitive. This process dictates to the system the maximization of the expected success of a grasp and the minimization of the risk involved in the implementation process in the environment of partial observability. The RGB-D perception has evaluated the proposed framework in a simulated and in a real-life cluttered environment. Empirical findings show that uncertainties are indeed a major strength of grasp success and overall robustness as compared to deterministic baselines especially when faced with a dense clutter and huge occlusions. The implications of these findings are that uncertainty-conscious perception is essential to the stable manipulation of objects in complex real-world environments, and indicates a promising future of the realization of safer and more autonomous grasping systems.