DeepMind Lab#

TorchWM supports DeepMind Lab through DMLabEnv, a lightweight adapter around the native deepmind_lab Python module. It converts Lab RGB observations into the channel-first image dictionary used by Dreamer and other pixel-based world-model code.

Install: pip install dmlab-gym && dmlab-gym build (or install deepmind_lab directly)

Dreamer uses cfg.env_backend = "dmlab" to select this backend. DMLab-specific fields (dmlab_action_repeat, dmlab_observations, dmlab_action_set, dmlab_config, dmlab_renderer) are forwarded from DreamerConfig to DMLabEnv. See Dreamer: Model-Based RL with Latent Dynamics for the full Dreamer config reference.

Direct usage#

from torchwm import make_dmlab_env

env = make_dmlab_env("rooms_collect_good_objects_train", seed=0, size=(64, 64))
obs = env.reset()
action = env.action_space.sample()
obs, reward, done, info = env.step(action)

The default action space is a normalized one-hot Box[-1, 1] over a compact set of navigation actions. You can pass action_set= to DMLabEnv or set cfg.dmlab_action_set to use a custom list of native seven-element Lab actions.

Backend options#

Option	Default	Description
`cfg.dmlab_action_repeat`	`4`	Native Lab frame repeat passed to `env.step(..., num_steps=...)`.
`cfg.dmlab_action_set`	`None`	Optional custom 2D array of native Lab actions.
`cfg.dmlab_observations`	`None`	Additional native observation names. `RGB_INTERLEAVED` is always included.
`cfg.dmlab_config`	`None`	Extra Lab config values. Width and height are derived from `cfg.image_size`.
`cfg.dmlab_renderer`	`"hardware"`	Renderer argument forwarded to `deepmind_lab.Lab`.

TorchWM’s shared Dreamer wrapper stack still applies cfg.action_repeat outside the DMLab adapter. If you only want native Lab frame repeat, leave cfg.action_repeat = 1.