Environments

An environment owns the scene, robots, observations, and restorable world state. Tasks can reuse one environment with different tools, instructions, and success criteria.

The simulation classes documented here use ManiSkill. DefaultEnv supplies MAGMA integration, while your subclass defines scene loading, episode initialization, and extra observations. DefaultMultiAgentEnv supports the multi-robot path; it is an additional implementation choice, not a requirement for a first scenario.

Initialization and transitions

InitializationParameters.env_options supplies episode-reset options. A stage's entry_transition can modify selected environment slots before that stage starts. reset_at_end restores the task's reference environment state at a stage boundary. Use a reset only when the next interaction should not preserve the preceding physical changes.

Restorable state

Simulation snapshots must also preserve non-physics state that affects future behavior: machine modes, dirt flags, counters, or occupancy. get_magma_extra_state and set_magma_extra_state integrate such data with DefaultEnv.get_state_dict and set_state_dict under magma_extra_state.

Observations

Expose the data tools and goals need through _get_obs_extra. Document names, shapes, and meanings. Some existing scenarios expose pose tensors directly; others use structured entries. These are scenario contracts, not interchangeable global formats.

Keep environment batches independent. A tool targets one env_id; goals verify a batch; transitions receive selected env_ids.

Next: Create an environment or add transitions and extra state.