24. March 2021 : New publication on embodied learning research. Training an agent in VR by deep reinforcement learning.
Clay V, König P, Kühnberger KU and Pipa G (2020)
Learning sparse and meaningful representations through embodiment
Neural Netw 134:23-41
How do humans acquire a meaningful understanding of the world with little to no supervision or semantic labels provided by the environment?
To answer this question, we investigated the influence of embodiment by means of a deep reinforcement learning agent that was trained before in a 3D environment with very sparse rewards.
Our results show that the agent not only learnt to reliably represent the action relevant information extracted from a simulated camera stream, without receiving any semantic labels. Even more so, the quality of the representations learnt suggest that embodied learning is more efficient than fully supervised approaches.