This repository collects two experiments conducted to explore the effect of environment distractions on the representations learned by an AE and the proposed Encoder-Decoder structure (RAED). In both experiments we use the AURORA [1] algorithm to learn robot behaviours.
airhockey
simulates an air-hockey table on which a multi-joint planar robot arm is fixed. The robot tries to explore its capability to create distinct air-hockey puck trajectories by interacting with the puck.
The environment distractions are modelled by the appearance of an additional puck that spawns and moves according to a random force vector.
objectarrangement
simulates a room in which the same robot arm instead tries to move a heavy object to different positions in a room. Environment distractions materialise in the appearance of additional objects.
[1] Antoine Cully. Autonomous skill discovery with Quality-Diversity and Unsu-pervised Descriptors. 9, 2019. pages 1, 2, 5, 23
For a detailled analysis of the experiment results, please refer to the PDF in the repository.