Autonomous Learning of Object-specific Grasp Affordance Densities

In this paper, we address the issue of learning and representing object grasp affordances. Our first aim is to organize and memorize, independently of grasp information sources, the whole knowledge that an agent has about the grasping of an object, in order to facilitate reasoning on grasping solutions and their likelihood of success. By grasp affordance, we refer to the the different ways to place a hand or a gripper near an object so that closing the gripper will produce a stable grip. The grasps we consider are parametrized by a 6D gripper pose and a grasp (preshape) type. The gripper pose is composed of a 3D position and a 3D orientation, defined within an object-relative reference frame. We represent the affordance of an object for a certain grasp type through a continuous probability density function defined on the 6D object-relative gripper pose space SE(3), similar to the approach of de Granville et al. [2]. The computational encoding is nonparametric: A density is simply represented by the samples we see from it. The samples supporting a density are called particles; the probabilistic density in a region of space is given by the local density of the particles in that region. The underlying continuous density is accessed by assigning a kernel function to each particle – a technique generally known as kernel density estimation [6]. The kernel functions essentially capture Gaussian-like shapes on the 6D pose space SE(3) (see Fig. 1). The expressiveness of a single kernel is rather limited: location and orientation components are both isotropic, and within a kernel, location and orientation are modeled independently. Nonparametric methods account for the simplicity of individual kernels by employing a large number of them: a grasp density will typically be supported by a thousand particles. An object is linked to a separate grasp density for each type of grasp it affords, e.g. one density for pinch grasp affordance and another density for or power grasps. The second contribution of this paper is a framework that allows an agent to learn initial affordances from various grasp cues, and enrich its grasping knowledge through experience. Affordances are initially constructed from human imitation, or from model-based methods [1]. The grasp data produced by these grasp sources is used to build continuous grasp hypothesis densities. Given the nonparametric representation, building a density from a set of grasps is straightforward – grasps can directly be used as particles representing the density. These densities are attached to a