Bayesian optimization for robust robotic grasping using a sensorized compliant hand

TL;DR

An experimental evaluation in the robotic system shows the usefulness of the use of Bayesian optimization techniques for performing unknown object grasping even in the presence of noise and uncertainty inherent to a real-world environment.

Abstract

One of the first tasks we learn as children is to grasp objects based on our tactile perception. Incorporating such skill in robots will enable multiple applications, such as increasing flexibility in industrial processes or providing assistance to people with physical disabilities. However, the difficulty lies in adapting the grasping strategies to a large variety of tasks and objects, which can often be unknown. The brute-force solution is to learn new grasps by trial and error, which is inefficient and ineffective. In contrast, Bayesian optimization applies active learning by adding information to the approximation of an optimal grasp. This paper proposes the use of Bayesian optimization techniques to safely perform robotic grasping. We analyze different grasp metrics to provide realistic grasp optimization in a real system including tactile sensors. An experimental evaluation in the robotic system shows the usefulness of the method for performing unknown object grasping even in the presence of noise and uncertainty inherent to a real-world environment.

Figures (9)

• Figure 1: Different setups used during experimentation with Bayesian optimization-based grasps: A) Simulation scenario, optimizing the grasp by moving CLASH hand freely around the object; B) Simulation of the complete robot model, with motion and collision constraints due to the arm and workbench; C) Real experimentation environment, including CLASH, a compliant under-actuated hand with tactile sensors on the fingertips.
• Figure 2: New metrics proposed: A) Approximation Reward in two grasps with collision, from higher $AR$ (left) to lower $AR$ (right). B) Contact Reward in two grasps with finger contact points, from higher $CR$ (left) to lower $CR$ (right).
• Figure 3: Objects from the YCB set used for experimentation, both in simulation and in the real environment. They show variety in shapes, weights, sizes, symmetry or the position of the center of mass.
• Figure 4: Comparison of the evolution of the $y_{simple}$ outcome, obtained using different implementations of the grasp evaluation function.
• Figure 5: Some of the best grasps obtained with each metric in simulation, grasping the mustard bottle, the mug, and the power drill. From left to right: $Q_{iso}$ ($w_1$ = 1), $Q_\epsilon$ ($w_2$ = 1), $Q_v$ ($w_3$ = 1) and $Q_{uni}$ ($w_4$ = 1).