EmblaV1

Extending DreamerV3 for better Morphology awareness by implementing a FFKSM

Details

A detailed description of EmblaV1 is provided at additionalMaterials/DreamerFFKSM.pdf.

Usage

• install requirements (note that you need torch with cuda)
• replace pusher_v5.xml or ant_v5.xml in the gymnasium lib folder (yourPythonVenv/gymnasium/envs/mujoco/assets) with the file in this repo! (necessary for the colourfilter)
• run main.py to start
• modify config to change all relevant Dreamer and FFKSM parameters(havent tested ant env with EmblaV1 yet)

Current state

• EmblaV1 done -> Result: DreamerV3 is already very good with limited vision and seems to create its own intrinsic selfmodel -> SM is redundant for EmblaV1

Acknowledgements

• NaturalDreamer Helped me to understand the DreamerV3 Architecture. Used the Code as a base.
• SelfModel Original Selfmodel-code, which I modified and fused with the DreamerV3 Architecture.