direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Inhalt des Dokuments

Life-long grasp success learning

Wednesday, 18. April 2018

Lupe

Masters

Motivation:

Grasping is a hard task for robots, because it is impossible to model the world and the robot dynamics perfectly. Without exhaustive modeling, we can increase the grasp performance of the robot by exploiting features of the environment like the wall of a crate. This features are called environment constraints (EC). The robot can push the object to one of the walls, by which the object motion is constrained along the wall and its location uncertainty is also reduced.

If there are multiple walls, the planner needs to decide which wall to use. Because modeling is not suitable for this task, the robot needs to learn from experience.

 

 

Description of Work

We implemented a planner that can detect ECs (surfaces, walls, and edges) and it generates a grasping graph (ECE graph). This graph encapsulates all the possible graps strategies that exploit an existing EC in the scene.  However, our planner is not yet enabled to learn which EC promises better grasp for a given object. Your task is to extend our planner such that after a grasp attempt it learns a bit more about the relation between environment constraints and object. So, the robot can learn from experience without the need of increased amount of data. 

For this project you need strong programming skills (python or C++). Experience with reinforcement learning is a plus.

 

People 
Előd Páll
Oliver Brock

 

back

Zusatzinformationen / Extras

Quick Access:

Schnellnavigation zur Seite über Nummerneingabe

Auxiliary Functions