Generic Planning Using Qualitative Causal Models for Self-Learning Autonomous Robots

Marco Van Cleemput, Jan Lemeire, Ruben Spolmink

Track: Type A (Regular Papers)

Keywords: Qualitative models, open-ended learning, autonomous robots, developmental learning

Abstract:

This paper puts forward qualitative causal models as the
basis for a cognitive architecture for autonomous agents that can learn
world models and use them for control and planning. The causal rela-
tions are described qualitatively by the direction of changes of the causal
influences. Context is taken into account when causal processes only take
effect for particular states. The model can be learned bottom-up based
on a few samples gathered during an exploration phase. During exploita-
tion, the model is used for planning and control to perform certain tasks.
The approach is demonstrated on a robot with a gripper that can move
objects and push buttons to open boxes. This approach is more sample
efficient than reinforcement learning, provides an explanation, is more
general and better reflects how humans learn and reason.

Serve As Reviewer: ~Jan_Lemeire1

Submission Number: 82