Intrinsically motivated RL agent for robotic manipulation tasks
Title: Intrinsically motivated RL agent for robotic manipulation tasks
SNIC Project: Berzelius-2022-22
Project Type: LiU Berzelius
Principal Investigator: Wenhao Lu <>
Affiliation: Chalmers tekniska högskola
Duration: 2022-01-31 – 2022-08-01
Classification: 10207


In environments where RL agent sparsely interacts with objects, it's encouraged to motivate the agent to intrinsically explore via a well-shaped intrinsic motivation-type objective. Our idea in this work is that learning can be driven by knowing what an object is used for (i.e., its affordance) and how it can be connected to other objects. To this end, we introduce a new measure of intrinsic objective that both encourages the agent to have maximum control over objects (whose affordance is under-explored) and one object to control the other. The expected empirical result for robotic manipulation tasks: multi-cube stacking would show strong effectiveness of this learning paradigm and that our new intrinsic motivation is able to aid few-shot learning for stacking more cubes without task rewarding, thus significantly reducing huge resource computation normally required for a range of real-world robotic tasks