Deep Reinforcement Learning for vision-based robotic grasping

Robotic grasping represents a major competency for the safe and effective handling of objects in unstructured settings. Stability in grasps, especially when considering objects whose shape or properties are irregular or non-uniform, still presents a challenge deriving principally from uncertainties inherent to contact geometry and friction properties and mass distribution. This thesis presents learning methods for the enhancement of the robustness of grasps through explicit consideration of the influence of the center of mass (CoM) on stable manipulation. We present a new system that integrates RGB-D perception, pose generation for grasps, and reinforcement learning to forecast the CoM and control stable grasp execution. Our approach integrates two candidate grasp sampling techniques---a major-axis method for single-axis objects and grid sampling procedure for multi-axis objects---with a tabular Q-learning algorithm that learns stable grasp settings by self-exploration in an iterative fashion. Unlike previous approaches highly reliant on tactile force sensors or CNNs, our approach parameters grasp instabilities explicitly through on board vision and leverages lightweight learning to correct CoM estimation and grasp policies. The system was evaluated and validated using a 6-DoF xArm6 robotic arm equipped with a parallel jaw gripper and onboard Intel RealSense cameras. A custom 3D-printed object set with changing mass distributions was used to verify performance for different CoM conditions. Experimental and simulation results demonstrate that the proposed method can effectively approximate the CoM for single-axis objects, recognize stable grasp pose even when CoM estimation is questionable, and increase the rate of successes in repeated grasps. For multi-axis objects, the method provides stable predictions along the dominant axis while maintaining tenable grasps for more challenging configurations. In short, this thesis provides a practical and interpretable learning-centered framework for grasp stability that only requires minimal tactile sensing, blends well with standard robotic perception and planning systems, and highlights the capacity of reinforcement learning to enhance robustness in ordinary robotic manipulation tasks.