Multi-Arm Bandit Models for 2 D Sample Based Grasp Planning with Uncertainty

For applications such as warehouse order fulfillment, robot grasps must be robust to uncertainty arising from sensing, shape, mechanics, and control. One way to achieve this is to evaluate the performance of candidate grasps by sampling perturbations in shape, pose, and control, computing the probability of force closure for each candidate to identify the grasp with the highest expected quality. Prior work has turned to cloud computing because evaluating the quality of each grasp is computationally demanding. To improve efficiency and extend this work, we consider how Multi-Armed Bandit (MAB) models for optimizing decisions can be applied in this context. We formulate robust grasp planning as a MAB problem and evaluate grasp planning convergence time using 100 object shapes randomly selected from the Brown Vision 2D Lab Dataset from 1000 uniformly distributed candidate grasp angles. We consider the case where shape uncertainty is represented as a Gaussian process implicit surface (GPIS) and there is Gaussian uncertainty in pose, gripper approach angle, and coefficient of friction. Uniform allocation and iterative pruning, the non-MAB methods, converge slowly. In contrast, Thompson Sampling and the Gittins index method, the MAB methods we consider, converged to within 3% of the optimal grasp 5x faster than iterative pruning.