Learning Generalizable Grasp Metrics
for Natural Grasping with High-dimensional Robotic Hands
Nathan Ratliff, J. Andrew Bagnell, Siddhartha Srinivasa
Grasp planning is a difficult problem. Robot hands are high-dimensional tools that must be maneuvered into a configuration that prevents the object from being removed. Grasp discovery requires understanding both the high-dimensional configuration space of the hand, as well as its relation to the geometrical characteristics of the object.
A well studied approach to grasp planning is to reduce it to discrete optimization in the tradition of the GraspIt! system (Miller et al., 2000 link). This approach chooses a discrete set of preshape configurations, each oriented toward the object and situated a short distance from the object with the fingers spread. For each preshape, a simple controller moves the hand forward and closes the fingers around the object to attempt a grasp. Each final grasp configuration is then scored and the optimizer returns the optimally scoring grasp. In essence, this approach is a straightforward technique for simply generating and scoring a collection of grasp candidates.
My work in grasp planning focuses on generalizing demonstrated grasp patterns to learn a score function that suits our needs. Many grasp metrics score grasp configurations using quantities derived from force closure. Force close measures the extent to which a grasp can withstand a force placed on the object from any direction; its allure follows from its ease of computation and its conservative nature. Unfortunately, force closure often results in strong, rigid grasps, and is therefore desirable only in select situations.
Our experiments show that we can successfully learn more general grasp metrics from expert grasp demonstrations that produce looser, more natural, grasps which often only cage the object. Our learned grasp metrics are functions of local features that quantify the geometrical relationship between the hand configuration and the object surface. Since these features can be extracted for any grasp candidate and any object, we can test the generalization performance of our approach on previous unseen objects. The top two rows of images show the prediction performance of our learned grasp metric on previously unseen objects alongside the desired grasp chosen by an expert teleoperator for the object. In all cases, we trained our metric to choose from a set of 2,496 distinct grasp configurations each positioned around the same approach direction. We relegate the choice of grasp point on the object to alternative techniques such as the image-based grasp point detection algorithm described in (Saxena et al. 2007). The lower collection of images demonstrate how grasp performance varies across the same object given differing approach directions and grasp points. The robot hand used for these experiments was the ten degree-of-freedom BarrettHand developed at Barrett Technology.