guo925658/visual-pointmaze-mctd-v1

Visual PointMaze — MCTD reimplementation v1

Complete Medium and Large local RGB datasets used for the visual auxiliary training in Looped Fast-MCTD. There are 2,202,200 frames in 1,076 NPZ shards, preserving the original train/validation episode split. The four ZIP files include the original state trajectories as well as every RGB shard and its checksum receipt.

This is our re-rendering of OGBench PointMaze state trajectories, using an agent-centered camera configured by this project. It is not an official Fast-MCTD visual dataset release or an exact reproduction of the authors' camera.

Environment	Split	Episodes	RGB frames	NPZ shards	Download
Medium	train	1,000	1,001,000	489	ZIP
Medium	validation	100	100,100	49	ZIP
Large	train	1,000	1,001,000	489	ZIP
Large	validation	100	100,100	49	ZIP

The compressed NPZ payload is 4,481,370,034 bytes in total. ZIP containers use storage without additional compression and include about 45 MB of source trajectories. Exact archive sizes and SHA256 values are in manifest.json and SHA256SUMS.

Download and read

Use Python with huggingface_hub and numpy installed. Set local_dir to a location with sufficient disk space; keeping the ZIPs and extracting both complete environments needs about 10 GB. For a pinned release, replace main with the commit ID from the dataset's commit history.

from pathlib import Path
from zipfile import ZipFile
import hashlib
import json
import numpy as np
from huggingface_hub import snapshot_download

root = Path(snapshot_download(
    repo_id="guo925658/visual-pointmaze-mctd-v1",
    repo_type="dataset",
    revision="main",
    local_dir="./visual-pointmaze-mctd-v1",
))
manifest = json.loads((root / "manifest.json").read_text())
for item in manifest["archives"]:
    path = root / item["path"]
    digest = hashlib.sha256()
    with path.open("rb") as stream:
        for chunk in iter(lambda: stream.read(4 * 1024 * 1024), b""):
            digest.update(chunk)
    assert digest.hexdigest() == item["sha256"], path
    with ZipFile(path) as archive:
        archive.extractall(root)

with np.load(root / "medium/full/train/00000000-00002048.npz",
             allow_pickle=False) as batch:
    print(batch["observations"].shape)  # (2048, 64, 64, 3), uint8
    print(batch["actions"].shape)       # (2048, 2), float32

These are native trajectory NPZ files in ZIP containers, not a tabular datasets.load_dataset export. You may download only the archives for the environment/split you need. Each ZIP extracts its split, the full environment manifest and identity, and the corresponding original state NPZ under source_states/. Shared manifest/identity files in the two splits are identical.

Data schema and episode boundaries

Field	Shape	dtype	Meaning
observations	N × 64 × 64 × 3	uint8	RGB at source state t, before action t
actions	N × 2	float32	Original source action at t
xy	N × 2	float32	Ground-truth position for offline labels/audits
terminals	N	bool	End of the recorded trajectory, not task success
source_indices	N	int64	Contiguous zero-based index within the original split

Each source episode has 1,001 records for a 1,000-step trajectory. Preserve the terminal record; do not interpret it as an additional valid within-episode transition. Do not form next-state pairs or three-frame inverse-dynamics inputs across a terminal. Shards hold at most 2,048 frames and may cross episode boundaries; an NPZ file is not an episode. The source indices and terminal flags allow reconstructing complete trajectories. No train/validation mixing or subsampling was performed.

The xy field is a supervised label, not an extra visual observation. The intended visual planner/controller consumes an 8D VAE latent from RGB; true position must not be concatenated to it. Evaluation goals are produced by the simulator, not stored as per-frame goal labels in these training shards.

Camera and provenance

• PointMaze Medium/Large, 64 × 64 RGB, original colored coordinate-gradient floor.
• Camera tracks agent XY with fixed orientation: distance 10, azimuth 90 degrees, elevation -60 degrees, look-at height 0.7; no goal marker.
• Source training files: pointmaze-{medium,large}-navigate-v0.npz; validation files: pointmaze-{medium,large}-navigate-v0-val.npz.
• Rendering recipe SHA256: 4f01638092c9b39523286644ffe0fd26525aa2f72f6aaeaa3fdbc6f9b28097b6.
• Historical simulator transition/camera preflight SHA256: 9180a2bcc50e69075209ba94582f77f725097422471c05c7329ca69664435ae4.
• The historical preflight recipe and full rendering recipe have different hashes; both are retained unchanged. Per-shard receipts identify the recipe that actually generated the data.

reproduction/ contains pinned configuration files, renderer/preflight/export code, sample images and historical provenance. The simulator is the customized OGBench wheel bundled in the linked GitHub repository, not an assumed equivalent current PyPI version. The original rendering scripts use the recorded Palmetto Slurm and /scratch/gguo/idea-loop-mctd paths; moving to another machine requires adapting paths and rerunning its simulator preflight. Reading the released NPZ files requires neither Slurm nor a simulator.

Validation and experiment status

Before release, all 1,076 shard SHA256 hashes were checked and every array in every NPZ was decoded. Shapes/dtypes, contiguous source coverage, episode boundaries, and exact action/XY/terminal correspondence to the original hashed state trajectories were checked. ZIP member CRCs were checked after packaging. These are offline data integrity checks, not new benchmark evaluations.

The two datasets were generated on 2026-09-15. VAE, inverse-dynamics and position auxiliaries completed 50,000 updates per environment, with evidence audited on 2026-09-16. The visual Fast-MCTD/Loop planners and final success-rate evaluations have not been completed. Auxiliary weights are not part of this data release. The sparse-subgoal controller bridge and visual guidance integration remain unfinished; auxiliary validation errors are not success rates.

The authoritative experiment results remain in the project report. Any future performance evaluation must follow its fixed paper contract: five task IDs × ten seeds per environment/method/planning mode, Replan interval 50, paired initial states, identical controller assets and external search budget, and separately recorded Loop work and wall time.

Attribution

Source trajectories and simulator: OGBench, Seohong Park, Kevin Frans, Benjamin Eysenbach and Sergey Levine, ICLR 2025. The bundled OGBench MIT notice is preserved as LICENSE-OGBench; project code licensing is preserved as LICENSE. The re-rendered views and selected camera configuration are from the Looped Fast-MCTD project.

Method references: Fast-MCTD and original MCTD. The source trajectory generator is the authors' pinned OGBench fork.