Assistive 6-DOF Robotic Grasping of Unknown Objects

Picking up an unknown object is a central problem in robotics that has not yet been fully solved. While many approaches operate in a controllable environment, such as the industrial one, we target the assistive context. Instead of a known and static workspace with clearly defined action constraints, our considered environment is unstructured and unknown. A large number of common, implicit assumptions cannot be made. The initial sensor measurements of imaging sensors might not show a valid grasp solution, motion goals of the manipulator might not be reachable or are in collision and the sensor measurements do not necessarily have to be within the domain of the training data. In addition, the assistive context offers a variety of unique sensory challenges. The scene might be partially over-illuminated and the surface properties can vary greatly, which implies that measurement noise is not identically distributed. These challenges cannot be overcome by a pre-configured, optimal sensor positioning or software-setting, but rather must be dealt with as is during runtime. We propose a closed-loop, end-to-end algorithmic pipeline that deals with assistive, robotic grasping in a holistic manner. A probabilistic sensor fusion method is introduced that accurately reconstructs the desired scene segment and enables us to reliably estimate the local reconstruction accuracy. Furthermore, we present an exploration and motion strategy that allows the robot to autonomously search for grasp options while ensuring user safety at controller level. Finally, we propose a probabilistic grasp sampling and evaluation algorithm that does not rely on prior object knowledge or a dataset tailored to the use case but instead utilizes the local estimation uncertainty to generate robust grasps in an online fashion. The pipeline is real-time capable and extensive simulated as well as real experiments demonstrate that our method is robust and accurate, even with challenging household objects.