Active Learning Of Manipulation Sequences

Nicola Covallero, David Martínez, Guillem Alenyà, Carme Torras


Manipulation planning of cluttered objects involves a mixture of symbolic and geometric constraints which makes such planning very time consuming and often unsuitable for real applications. We propose to divide the geometric restrictions in two groups. The ones in the first group are used to generate a set of symbolic states used for planning. The evaluation of the ones in the second group is delayed after planning, and only relevant ones are evaluated when necessary. Each evaluation generates new states, and replanning is performed when necessary. The proposed system is able to get in few seconds a suitable sequence of pushing and grasping actions to achieve the goal of clearing a table. Its main advantage is that plans at a deterministic symbolic level with limited geometrical information, making the planning stage fast. We present a simple but effective implementation using only two actions and four pushing directions, and checking only the feasibility of the first action in the plan. Three different experiments validate the proposed system and show its potential in real robotic scenarios.


When the object is free of colision it is grasped.

If there is a collision, a push action in the convenient direction moves the object to a suitable position.

The length of the translation depends on the neighbor objects.

If the robot hand is not well modelled, push actions can cause colisions with other objects