Haptic Object Identification for Advanced Manipulation Skills
129
This approach, known as tactile and haptic exploration, enables robots to significantly increase and extend the results of visual object identification methods.
In contrast to recent research in haptics, this work presents an online inference
algorithm which is capable of acquiring information not only about the geometry
but also about the material parameters of an unknown object.
The remainder of this work is structured as follows: the next section outlines
the mathematical problem tackled in this work, followed by an outline on how
this work is positioned compared to related work in Sect. 2. The concepts of the
proposed grid-based and the shape-based exploration strategies are sketched in
Sect. 3. The idea of classifying individual components by their material types is
shown in Sect. 4, whereas Sect. 5 outlines the evaluation of the proposed methods
in a simulated environment. The summary in Sect. 6 concludes this work.
1 Problem Statement
The task of haptic object identification consists of two main challenges. First,
the geometric shape of an object, denoted as M in the context of this work, is
in general unknown. Second, the object is characterized by an undefined paramterization Θ, that describes the material properties of an object, e.g. a material
classificator that maps each component of M to a finite set of materials. Given
the state, control inputs and measurements of a robot as x 1..t = {x 1 , . . . , x t },
u = {u 1 . . . u t } and r 1..t = {r 1 , . . . , r t } from time-step 0 to the current time
step t, the problem is given by finding proper mapping functions
M ← F m (x 1.. t, u 1.. t, r 1.. t, Θ),
(1)
Θ ← F p (x 1.. t, u 1.. t, r 1.. t, M ).
(2)
Finding proper mappings F m and F p is non-trivial as they are in general dependent on each other. Nonetheless, when analyzing the problem individually, one
can relax these problems and focus on finding these mappings for a fixed Θ or
M . As a variety of promising methods on solving these problems individually in
literature exists, we continue with an overview of related work.
2 Related Work
Although vision has been established as the backbone of robotic perception,
haptic information acquisition has been used to understand or recognize shapes
of objects for years [1]. Navarro et al. [16] present an approach for haptic object
recognition based on extracting key features of tactile and kinesthetic data using
a clustering algorithm, where a tactile sensor performs haptic sensation tasks
using different robotic hands. Behbahani et al. [3] have introduced haptic Simultaneous Localisation and Mapping (SLAM) into the field of haptic exploration,
which is inspired by visual SLAM techniques [6] and occupancy grid methods [7].
Through adaption of the FastSLAM [15] algorithm, a novel method is proposed
to iteratively learn the shape of the surface of objects. The same approach is
129
This approach, known as tactile and haptic exploration, enables robots to significantly increase and extend the results of visual object identification methods.
In contrast to recent research in haptics, this work presents an online inference
algorithm which is capable of acquiring information not only about the geometry
but also about the material parameters of an unknown object.
The remainder of this work is structured as follows: the next section outlines
the mathematical problem tackled in this work, followed by an outline on how
this work is positioned compared to related work in Sect. 2. The concepts of the
proposed grid-based and the shape-based exploration strategies are sketched in
Sect. 3. The idea of classifying individual components by their material types is
shown in Sect. 4, whereas Sect. 5 outlines the evaluation of the proposed methods
in a simulated environment. The summary in Sect. 6 concludes this work.
1 Problem Statement
The task of haptic object identification consists of two main challenges. First,
the geometric shape of an object, denoted as M in the context of this work, is
in general unknown. Second, the object is characterized by an undefined paramterization Θ, that describes the material properties of an object, e.g. a material
classificator that maps each component of M to a finite set of materials. Given
the state, control inputs and measurements of a robot as x 1..t = {x 1 , . . . , x t },
u = {u 1 . . . u t } and r 1..t = {r 1 , . . . , r t } from time-step 0 to the current time
step t, the problem is given by finding proper mapping functions
M ← F m (x 1.. t, u 1.. t, r 1.. t, Θ),
(1)
Θ ← F p (x 1.. t, u 1.. t, r 1.. t, M ).
(2)
Finding proper mappings F m and F p is non-trivial as they are in general dependent on each other. Nonetheless, when analyzing the problem individually, one
can relax these problems and focus on finding these mappings for a fixed Θ or
M . As a variety of promising methods on solving these problems individually in
literature exists, we continue with an overview of related work.
2 Related Work
Although vision has been established as the backbone of robotic perception,
haptic information acquisition has been used to understand or recognize shapes
of objects for years [1]. Navarro et al. [16] present an approach for haptic object
recognition based on extracting key features of tactile and kinesthetic data using
a clustering algorithm, where a tactile sensor performs haptic sensation tasks
using different robotic hands. Behbahani et al. [3] have introduced haptic Simultaneous Localisation and Mapping (SLAM) into the field of haptic exploration,
which is inspired by visual SLAM techniques [6] and occupancy grid methods [7].
Through adaption of the FastSLAM [15] algorithm, a novel method is proposed
to iteratively learn the shape of the surface of objects. The same approach is