242
have implemented the following sequence: f) — c) — a) — b) — d) — e).
The calculation of each of these features returns a float value between 0.0 and 1.0, where 0.0 means that the
feature is absent and 1.0 that the feature is present. These values are then used to assess the usability of the
vertex pair as a grasping opportunity. The result should be consistent with the feature-values, e.g., if one feature
is 0.0, the result should be 0.0, too. Fuzzy set theory provides several functions for the conjunctive combination
of evidence, the so-called triangular norms, or T-norms which have the appropriate qualitative effect, among
them T; (a, b) = min(a, b) or T(a, b) — a x b. Using the latter, our final grasping opportunity measure for one
vertex pair becomes:
Ti = Uf fi * UcCi * Vals * Ub; * vadi * veei
where 7; stands for “ranking of vertex pair i” and v, denotes the relative importance of feature x, with
x € {a,b,c,d,e, f}. Finally, we select the vertex pair with the highest ranking and the corresponding grasp is
carried out, i.e., the object concerned by the grasp is removed.
5 VISION SYSTEM INTEGRATION
A state diagram of the complete vision system is shown in Figure 3. A short description of the tasks performed
in the different states follows:
- Start: The vision system and the robot are initialized.
- Image acquisition: The two depth images are taken and merged (as described in section 3.1 and 3.2)
- Build world representation: The input image is segmented and "objects" are assigned to each con-
nected surface region (see section 3.3). Several predicates are assigned to each object, such as "isolated"
or “on top of”. All objects are also marked "graspable".
- Finding grasping opportunities: As described in section 4, this state consists of an “object selection
sub-state” and a state which calculates the actual grasping points.
- Grasp object: The grasp is carried out by the robot. Independently of the success of the grasp, a new
image is taken.
- Mark object *non-graspable": If there are no suitable grasping opportunities for an object, it is
marked “non-graspable” and thus is not considered anymore by the "object selection state".
- Push highest object: If there are still objects in the scene but all of them are marked "non-graspable",.
the vision system has failed. The reason might be a high degree of occlusions, reflecting or highly textured
objects etc. The system now tries to change the spatial constellation of the scene. This is done in the
hope that the deadlock is removed.
©
Wait for Image acquisition
next scene
: Build world
Push highest representation
object
Figure 3: State diagram of the
Scene empty | Object selection | complete vision system
all objects : " Mark object =
marked "non-graspable" object selected non-graspable
Gens point grasping not possible
etermination
Finding
grasping
opportunities
Grasp object =
IAPRS, Vol. 30, Part 5W1, ISPRS Intercommission Workshop "From Pixels to Sequences", Zurich, March 22-24 1995
