Procgen

The Procgen backend adapts the procgen.ProcgenEnv vector API to TorchWM’s single-environment image interface for procedurally generated benchmark games such as CoinRun, Maze, Heist, and StarPilot.

Install: pip install torchwm[procgen]

Main APIs

from world_models.envs.procgen_env import make_procgen_env, list_procgen_envs

env = make_procgen_env("coinrun", seed=0, size=(64, 64))
obs = env.reset()
print(obs["image"].shape)  # (3, 64, 64)

The factory accepts Procgen shorthand names and Gym-style ids:

make_procgen_env("coinrun")
make_procgen_env("procgen-coinrun-v0")
make_procgen_env("procgen:procgen-coinrun-v0")

Use list_procgen_envs() to inspect the supported game names.

Dreamer uses cfg.env_backend = "procgen" (or "coinrun") to select this backend. See Dreamer: Model-Based RL with Latent Dynamics for the full Dreamer config reference.

Observations and actions

ProcgenImageEnv unwraps the leading vector dimension from ProcgenEnv(num_envs=1) and returns:

{"image": uint8 array with shape (3, H, W)}

Procgen actions are discrete. For consistency with TorchWM’s other discrete image adapters, the wrapper exposes a continuous one-hot-like action space with shape (n,) and values in [-1, 1]. The selected action is the index of the largest value in the model action vector.

Example script

TorchWM includes a compact Dreamer example for Procgen environments:

python examples/run_dreamer_procgen.py --env coinrun --total-steps 2000
python examples/run_dreamer_procgen.py --list-envs

Available games

TorchWM recognizes the standard Procgen games: bigfish, bossfight, caveflyer, chaser, climber, coinrun, dodgeball, fruitbot, heist, jumper, leaper, maze, miner, ninja, plunder, and starpilot.