zhangdw/minigrid

MiniGrid DoorKey provides frozen initial states for MiniGrid-DoorKey-8x8-v0. An agent must find a key, unlock a door, and reach the goal within 64 actions. Exact state restoration and fixed evaluation membership make this a useful setting for visual agentic reinforcement learning and credit assignment research.

2,886 Training states

256 Validation states

448 × 448 Native RGB observation

Quick start · Frozen splits · Interaction · State format · Reproducibility

Evaluation scope: this is a custom initial-state holdout split. Training and validation may share wall layouts, so results measure initial-state generalization. The split is independently constructed and is not an official Farama evaluation protocol.

Quick start

Download the package and run its offline restoration verifier in a dedicated environment:

uvx hf download zhangdw/minigrid --type dataset --local-dir minigrid
cd minigrid
uv venv --python 3.11 .uv-venv/minigrid
uv pip install --python .uv-venv/minigrid/bin/python -r requirements.txt
SDL_VIDEODRIVER=dummy .uv-venv/minigrid/bin/python code/restore_and_verify.py

The verifier checks file hashes, restores every training and validation state, compares the exact RGB pixels, and verifies split separation. It does not load a policy or model. Dependencies are pinned to MiniGrid 3.0.0, Gymnasium 1.2.3, NumPy 1.26.4, Pygame 2.6.1, and Pillow 12.3.0.

Frozen splits

File	States	Role
fairness/train.jsonl	2,886	Complete eligible training pool after validation groups are excluded
fairness/evaluation.jsonl	256	Canonical validation set, sorted by task ID
fairness/candidates.jsonl	3,343	Deduplicated candidate archive for audit

The source generator uses seeds 0–4095: 4,096 states yield 753 duplicates and 3,343 eligible candidates, with zero candidates unreachable within 64 actions. After selecting 256 validation states, 201 additional orientation variants are reserved with their validation groups. Thus 2,886 + 256 + 201 = 3,343.

Both task overlap and orientation-invariant split-group overlap are zero. Only fairness/evaluation.jsonl is the validation set. Selection uses environment structure and native shortest paths, with no model calls or model-performance filtering.

Interaction protocol

Property	Definition
Environment	MiniGrid-DoorKey-8x8-v0
Mission	use the key to open the door and then get to the goal
Action budget	64 native actions
Actions	left, right, forward, pickup, toggle
Observation	Full 448 × 448 RGB image, RGBImgObsWrapper(tile_size=56), highlight=False
Goal reward	1 - 0.9 × t / 64 on reaching the goal at action t; zero otherwise
Protocol	minigrid-doorkey-8x8-h64-v1

The key–door–goal sequence supports studying how earlier actions contribute to a later reward. The stored shortest paths are audit material, not step-level credit labels or policy inputs. Keep observations, reward, action budget, and fixed task membership consistent across methods.

Action and observation details

Native action IDs are left=0, right=1, forward=2, pickup=3, and toggle=5. Left and right rotate the agent; drop and done are not selectable policy actions. In the reference interaction protocol, an invalid response executes an internal native done/no-op, consumes one action, and preserves physical state. A supported action with no visible effect is still valid. Goal success on action 64 takes precedence when termination and truncation are both true. The environment adds no training penalty.

The RGB observation has no recoloring, coordinates, prior images, or reference paths. The reference policy context includes the mission, direction, holding status, remaining budget, and the last two compact action/feedback records. The prompt version is minigrid-vlm-v1; model and processor weights are supplied by the experiment.

State format

Each JSONL row contains initial_state and metadata. Initial states store the entire board and agent configuration, allowing exact restoration without procedural regeneration.

Field	Meaning
initial_state.grid	Native [object_type_id, color_id, object_state] cells, indexed as grid[x][y]
initial_state.agent_pos, agent_dir	Native (x, y) position and facing direction
initial_state.carrying	Encoded held key, or null
initial_state.step_count, max_steps	Elapsed actions and action budget; initially 0 and 64
metadata.task_id	SHA-256 of the complete initial state
metadata.split_group_id	State hash omitting only the initial facing direction
metadata.image_sha256	Hash of the restored raw RGB pixels
metadata.shortest_length	Verified shortest path length in native actions

Exact encoding and task identity

The state schema is minigrid-semantic-state-v1. Its exact keys are schema, env_id, mission, width, height, max_steps, step_count, grid, agent_pos, agent_dir, and carrying.

The grid is x-major, preserving empty cells, walls, the goal, key, and door. Door states are 0=open, 1=closed/unlocked, and 2=locked. Directions are 0=east, 1=south, 2=west, and 3=north. Inventory is null or the encoded yellow key.

Hashes use UTF-8 JSON with sorted keys, compact separators, ensure_ascii=False, and no added newline. The image hash covers raw C-order uint8 RGB pixels. Native actions are deterministic, so generator RNG state is unnecessary for restoration.

For matching decision states, the reference helper uses the task ID plus the full current semantic state, including elapsed steps; actor history is excluded. Object coordinates and stored solutions remain training/audit metadata and are not exposed in policy text.

Reproducibility

How the 256 validation states are selected

Selection version: minigrid-eval-256-v1. Native breadth-first search expands turns, forward movement, pickup, and toggle under the 64-action budget. Eligible candidates are restored, checked against their RGB hashes, and their shortest paths are replayed in the native environment. Seed provenance and rejections are recorded in fairness/provenance.json.

Shortest-length quartile boundaries are [14, 17, 19], using the nearest-rank rule. Equal lengths stay together, and empty bins are omitted. Length bins are crossed with splitting-wall x and door y. Hamilton's largest-remainder allocation assigns 256 slots proportionally; ties sort by (length_bin, wall_position, door_position).

Within each stratum, candidates rank by (SHA256(UTF8("minigrid-eval-256-v1:" + task_id)), task_id). Selected rows are sorted by task ID. Their complete orientation-invariant groups are excluded from training. Quotas, selected IDs, overlaps, and hashes are retained in fairness/evaluation-256.selection.json.

Additional verification and generation sources

To recompute native shortest paths for every training and validation state after the quick start:

SDL_VIDEODRIVER=dummy .uv-venv/minigrid/bin/python code/restore_and_verify.py --bfs

For complete generation and selection reproduction, use the matching generator implementation. Exact state and generation source snapshots are archived under code/. The archived preparation script imports that companion project's package; the standalone restoration verifier works directly from this dataset.

Preparation from existing frozen candidates repeats selection byte for byte. A per-seed generation journal supports resumption but is not part of the published release. For formal experiments, record the resolved Hub commit and pass it through --revision when downloading.

Path	Contents
code/restore_and_verify.py	Standalone file, state, image, and split verification
code/minigrid_fairness.py	Native restoration, validation, sampling, and BFS helpers
code/minigrid_protocol.py	Reference action grammar
fairness/evaluation-256.selection.json	Frozen selection audit
fairness/provenance.json	Candidate-generation provenance
fairness/reference-solutions.audit-only.json	Native action sequences for offline audit
checksums.json, fairness/checksums.json	Package and manifest hashes
requirements.txt	Pinned restoration dependencies

MiniGrid is developed by the Farama Foundation and distributed under the MIT license. This release contains synthetic DoorKey states. Dataset verification concerns state restoration and split integrity; training results must be measured separately.

Collection: Agentic RL & Credit Assignment

Related environments: ALFWorld · WebShop · Sokoban