Roadmap to v1.0

[Howuhh]

Roadmap

Approx v1.0 release: February 2024

General

• type hints
• tests
• documentation and examples
• environment registration
• colab notebooks with examples

Baselines

• PPO+RNN baselines (single task, meta-rl)
• open wandb reports + configs to reproduce paper figures
• multi-gpu training

Tiles

• Key (like in Minigrid)
• Door (like in Minigrid)
• Box (like in Minigrid) (may reduce FPS!!!)

Actions

• stochasticity (could be done with a wrapper)

Rules & Goals

• procedural generator (like in xland v2)
• pre-sampled benchmarks, 1M+ tasks

Map

• different grid layouts (mazes, rooms, objects)

Envs

• porting majority of minigrid envs
• full xland procedural meta-env
• observation/action spaces
• calibrate max_steps with available benchmarks and baselines
• provide rgb image observations

Beyond v1.0 directions

• stable env API (from gymnasium function_env ?)
• gymnasium adapters & wrappers
• multi-agent version (highly unlikely)
• MAML baselines, Mamba instead of GRU?
• Dataset + offline baselines (AD, DPT)
• procedural maze generation compatible with jit
• new INTERESTING rules and goals
• fast FOV algorithm for see_through_walls=False
• goals composition (OR, AND, NOT)
• masked rules as a part of the observation (could be done with a wrapper)
• rules descriptions with language
• porting multiroom, moving obstacles, safety envs from minigrid