Haptic Object Identification for Advanced Manipulation Skills
131
2.1 Contribution
In contrast to the stated work, this work outlines a haptic object identification
framework that allows to simultaneously refine the shape estimation and regress
the underlying material parameters as visualized in Fig. 1. In order to obtain F m ,
we incorporate findings from Haptic SLAM [3]. This encourages to maximize the
information gain at every step by applying concepts of Bayesian Filter theory [19]
in an iterative cycle of control and measurement updates. In order to decouple
the simultaneous parameter estimation problem, Θ is assumed to be fixed for
K e steps and only updated once collective data batches of update steps have
been obtained. In contrast to the cyclic nature of Bayesian Filters, this module
has access to a collection of data measures and can thus run nonlinear regression
techniques to regress the material parameters Θ. In the remainder of this work,
Θ describes the boundaries of a classifier that maps individual components of an
object to a set of available material types. Given this, F p is realized by applying
unsupervised clustering and model-fitting techniques.
1
3 Haptic Exploration
Before being able to extract information about the objects in the workspace, the
robot has to collect data through exploration. In order to gather this sensor data
in an efficient manner, we design a control flow for exploring our environment
based on a grid-based and a shape-based representation. Given initial data from
e.g. computer vision, an initial belief can be obtained, that can be iteratively
updated.
3.1 Grid-Based Exploration
We incorporate the findings from [3], where the belief of the geometry is stored
in an occupancy grid consisting of individual cells c ∈ C. We extend this to
M t = {M
0
t , M
1
t , . . . , M
Km
t
},
as an inference grid consisting of binary classifier layers M
k
t for K m material types, and an occupancy grid for k = 0, where each grid M
k
t assigns a
class-membership probability to each cell c. However, in contrast to storing
actual probability values in the grid as in [7], we use the log-odds-notation
P (c|r, x) =
exp M
k
t (c)
1+exp M k
t (c)
from [2] to store the current belief of each cell and
layer. With all layers being binary classifiers both measurements and states can
only take values in X = {0, 1}. As the haptic exploration seeks to maximize
the expected information gain, a utility metric needs to be defined that encourages to maximize information gain upon choosing the next cell to explore. We
incorporate findings from [12], that map the prior belief of a cell to all possible
1 The framework outlined in this work is not restricted to the presented identification
method. Nonetheless, this specific example serves as a proof of concept.
131
2.1 Contribution
In contrast to the stated work, this work outlines a haptic object identification
framework that allows to simultaneously refine the shape estimation and regress
the underlying material parameters as visualized in Fig. 1. In order to obtain F m ,
we incorporate findings from Haptic SLAM [3]. This encourages to maximize the
information gain at every step by applying concepts of Bayesian Filter theory [19]
in an iterative cycle of control and measurement updates. In order to decouple
the simultaneous parameter estimation problem, Θ is assumed to be fixed for
K e steps and only updated once collective data batches of update steps have
been obtained. In contrast to the cyclic nature of Bayesian Filters, this module
has access to a collection of data measures and can thus run nonlinear regression
techniques to regress the material parameters Θ. In the remainder of this work,
Θ describes the boundaries of a classifier that maps individual components of an
object to a set of available material types. Given this, F p is realized by applying
unsupervised clustering and model-fitting techniques.
1
3 Haptic Exploration
Before being able to extract information about the objects in the workspace, the
robot has to collect data through exploration. In order to gather this sensor data
in an efficient manner, we design a control flow for exploring our environment
based on a grid-based and a shape-based representation. Given initial data from
e.g. computer vision, an initial belief can be obtained, that can be iteratively
updated.
3.1 Grid-Based Exploration
We incorporate the findings from [3], where the belief of the geometry is stored
in an occupancy grid consisting of individual cells c ∈ C. We extend this to
M t = {M
0
t , M
1
t , . . . , M
Km
t
},
as an inference grid consisting of binary classifier layers M
k
t for K m material types, and an occupancy grid for k = 0, where each grid M
k
t assigns a
class-membership probability to each cell c. However, in contrast to storing
actual probability values in the grid as in [7], we use the log-odds-notation
P (c|r, x) =
exp M
k
t (c)
1+exp M k
t (c)
from [2] to store the current belief of each cell and
layer. With all layers being binary classifiers both measurements and states can
only take values in X = {0, 1}. As the haptic exploration seeks to maximize
the expected information gain, a utility metric needs to be defined that encourages to maximize information gain upon choosing the next cell to explore. We
incorporate findings from [12], that map the prior belief of a cell to all possible
1 The framework outlined in this work is not restricted to the presented identification
method. Nonetheless, this specific example serves as a proof of concept.