Statistical Identification of Composed Visual Features Indicating High Likelihood of Grasp Success

Mikkel Tang Thomsen, Leon Bodenhagen, Norbert Krüger

Publikation: Bidrag til tidsskriftKonferenceartikelForskningpeer review


In this paper, we address the problem of extracting
knowledge on the graspability of unknown objects based on
visual information using an Early Cognitive Vision system representing
contours and surfaces on different level of granularity.
We present an approach towards automatically identifying
configurations of three 3D surface features that predict grasping
actions with a high success probability. The strategy is based
on first computing spatial relations between visual entities and
secondly, exploring the cross-space of these relational feature
space and grasping actions. The data foundation for identifying
such indicative feature constellations is generated in a simulated
environment wherein visual features are extracted and a large
amount of grasping actions are evaluated through dynamic
Based on the identified feature constellations, we validate by
applying the acquired knowledge on a set of novel objects and
evaluating the proposed grasping actions in dynamic simulation.
In this validation, we have achieved an average success-rate
between 0.74 and 0.89 when relying on the information inherent
in configuration of three 3D surface features.


WorkshopBootstrapping Structural Knowledge from Sensory-motor Experience Workshop at IEEE International Conference on Robotics and Automation
Periode06/05/2013 → …