- RNN policies
- Extend configurability
- Allow recurrent baselines
- More RNN modules, incl Transformer
- Reward estimation extensions
- Auxiliary losses
- Curiosity
- Imitation learning
- Distributional perspective
- State/action modeling
- Sequence states/actions
- State-dependent actions
- Conditional/hierarchical actions
- Memory architecture
- Optimize retrieval of sequences
- Use TensorArray
- Improve other limitations
- CARLA environment
- Docs and assertions
- World's map loading (e.g. random, specific, etc.)
- Weather support
- Pretraining and Free play (e.g. for data collection)
- State space with a temporal component.
- ...
- To be determined...