Phantomare is still in the early stages and under heavy development.
Phantomare is a model-based RL algorithm originating from Dreamer-V3 with the goal of:
- Easy port to new problems
- Fast execution, training
- Clean code & design
- Good documentation
- conda -> venv
- Unit test suite (1D env for training quick tests?)
- Fix ReplayBuffer checkpoints
- Support discrete action spaces
- Verify performance on more baseline environments
- mypy typing
- Design front-end UI / API
- Support loading pretrained world-models
- Support training world-model on offline data
- algorithm speed ups (is torch scan available now?)
- Align algorithm with Dreamer V3 updates
- add evaluation cycle during training
- Support remote data collector
- Plan2Explore
- support training behavior on expert data
- Support PlaNet style behavior (planning/roll-out)
- explore alternative architectures (e.g., transformer-based)