Contextual Qualitative Deterministic Models for Self-learning Embodied Agents

This work presents an approach for embodied agents that have to learn models from the least amount of prior knowledge, solely based on knowing which actions can be performed and observing the state. Instead of relying on (often black-box) quantitative models, a qualitative forward model is learned that finds the relations among the variables, the contextual relations, and the qualitative influence. We assume qualitative determinism and monotonicity, assumptions motivated by human learning. A learning and exploitation algorithm is designed and demonstrated on a robot with a gripper. The robot can grab an object and move it to another location, without predefined knowledge of how to move, grab or displace objects.