Haptic Object Identification for Advanced Manipulation Skills
135
and the number of current geometric primitives is finite. This results in a deterministic classification, meaning that each measurement is assigned with a class
label. Having obtained these individual components, additional model-fits per
cluster result in new geometric primitive samples. By finally combining these
geometric primitives into a compositioned object a new particle is added to the
filter.
5 Evaluation
The outlined algorithm is evaluated with an artificial robot explorer using a
simulated environment using the Multi-Joint Dynamics with Contact (MuJoCo)
physics engine. The robot is equipped with a force-torque sensor that allows to
measure the impact during collision. In order to focus on the exploration process,
we directly explore and control in Cartesian space. Thus, the pose of the robot
x is controlled via a Cartesian impedance controller:
τ c = K s (x d − x) − K d ˙
x + τ ext ,
(8)
where τ c is the applied wrench command, x d is the desired pose, ˙
x the velocity,
K s and K d describe the stiffness and damping matrices, and τ ext describes an
additional feed-forward wrench command.
In order to test our methods against the challenges stated in Sect. 1, the robot
is faced with a set of unknown objects, which are composed of sub-components
of different materials, which differ in their stiffness values. MuJoCo [21] handles
all contacts between objects as soft constraints in the dynamic system, which
can be seen as a spring-damper system, where one can set the stiffness k m and
damping b m . These stiffness-damping values are artificial contact values used
for simulation dynamics rather than physically realistic values,
2 such as Young’s
modulus, that a robot can regress by obtaining measurements r t = (∗)x t , F t , Δl t ,
i.e. the magnitudes of force and displacement during contact at x t . In order
to assess the relationship between stiffness parameter k m of a MuJoCo object
model and the physical stiffness value k =
F
Δl , we fixed the damping values to
b m = 1 for all simulations and performed several experiments with increasing
parameter k m and compared the resulting estimations with the numeric stiffness
value k =
Fmax
Δlmax . Applying linear regression, the material stiffness is obtained as
k = f k (k m ) = γk m + δ ≈ (0.13 k m + 152) N /m.
(9)
Even though the data is just an approximation of the actual material data, it
is sufficient to evaluate the capability of our methods to differentiate between
materials and thus to identify the decomposition of an object. The unknown
exemplary objects are given as a composition of two boxes for object A and a
composition of two cylinders for object B, such that K m = 2. Their shape is
visualized as a ground-truth in Fig. 4, where the material classes are visualized
2 We refer to [20] for detailed information.
135
and the number of current geometric primitives is finite. This results in a deterministic classification, meaning that each measurement is assigned with a class
label. Having obtained these individual components, additional model-fits per
cluster result in new geometric primitive samples. By finally combining these
geometric primitives into a compositioned object a new particle is added to the
filter.
5 Evaluation
The outlined algorithm is evaluated with an artificial robot explorer using a
simulated environment using the Multi-Joint Dynamics with Contact (MuJoCo)
physics engine. The robot is equipped with a force-torque sensor that allows to
measure the impact during collision. In order to focus on the exploration process,
we directly explore and control in Cartesian space. Thus, the pose of the robot
x is controlled via a Cartesian impedance controller:
τ c = K s (x d − x) − K d ˙
x + τ ext ,
(8)
where τ c is the applied wrench command, x d is the desired pose, ˙
x the velocity,
K s and K d describe the stiffness and damping matrices, and τ ext describes an
additional feed-forward wrench command.
In order to test our methods against the challenges stated in Sect. 1, the robot
is faced with a set of unknown objects, which are composed of sub-components
of different materials, which differ in their stiffness values. MuJoCo [21] handles
all contacts between objects as soft constraints in the dynamic system, which
can be seen as a spring-damper system, where one can set the stiffness k m and
damping b m . These stiffness-damping values are artificial contact values used
for simulation dynamics rather than physically realistic values,
2 such as Young’s
modulus, that a robot can regress by obtaining measurements r t = (∗)x t , F t , Δl t ,
i.e. the magnitudes of force and displacement during contact at x t . In order
to assess the relationship between stiffness parameter k m of a MuJoCo object
model and the physical stiffness value k =
F
Δl , we fixed the damping values to
b m = 1 for all simulations and performed several experiments with increasing
parameter k m and compared the resulting estimations with the numeric stiffness
value k =
Fmax
Δlmax . Applying linear regression, the material stiffness is obtained as
k = f k (k m ) = γk m + δ ≈ (0.13 k m + 152) N /m.
(9)
Even though the data is just an approximation of the actual material data, it
is sufficient to evaluate the capability of our methods to differentiate between
materials and thus to identify the decomposition of an object. The unknown
exemplary objects are given as a composition of two boxes for object A and a
composition of two cylinders for object B, such that K m = 2. Their shape is
visualized as a ground-truth in Fig. 4, where the material classes are visualized
2 We refer to [20] for detailed information.