MAP-Data

📄 Paper  |  🌐 Project  |  💻 Code  |  📊 MAP-Bench  |  🤖 Models

MAP-Data contains forward RGB videos, task instructions, and frame-aligned annotations for navigation toward a Manipulation Anchor Pose (MAP), the base position and heading from which a robot can begin a manipulation. This public HM3D-derived subset excludes target instances present in MAP-Bench.

Trajectories	RGB frames	Hours at 15 Hz	Target instances
100,052	27,405,889	507.52	820

The download contains 10 independent ZIP shards (70.96 GB). RGB videos are 566 × 320 at 15 Hz.

Grounding across motion

Each row follows one released trajectory toward an oven (episode_090152), nightstand (episode_029224), or trash can (episode_112531). The nightstand route spans 4.7 m and changes heading by about 38 degrees, illustrating consistent target and approach-point annotations across a longer, oblique approach. The local orthographic view shows the forward-camera path in teal, the target outlined in amber, and three numbered observation poses. The orange point marks the planned approach location on the floor and its projection in each RGB view; green boxes localize the target. The three frames in each row come from one moving camera. The orthographic map and target outline are illustrative HM3D renderings; world poses and scene identifiers are not included in the released files.

Navigation, VLM grounding, and spatial reasoning

The videos, five alternative task instructions, target boxes, projected MAP points, and natural-language MAP descriptions support:

• VLM object grounding: pair a task instruction and RGB frame with the target bounding box to train referring-expression localization.
• Navigable-point grounding: use the projected MAP point and its natural-language location description to learn where the robot should approach and stop before interaction.
• Spatial and temporal reasoning: use frame-aligned boxes and MAP points along a moving-camera trajectory to construct supervision for instruction-conditioned spatial reasoning and consistency across changing viewpoints.

The annotations provide 2D spatial targets from successive views of one forward camera. For closed-loop navigation evaluation, use MAP-Bench.

Data format

Each of the 10 independent ZIP shards contains matching files under episodes/:

episodes/episode_000000.mp4
episodes/episode_000000.parquet
episodes/episode_000000.json

The JSON file contains only instructions (five task paraphrases) and map_description (a natural-language description of the MAP location). Sample an instruction when training, or select a fixed instruction and seed for evaluation. Anonymous episode IDs are stable and may have gaps.

Per-frame field	Meaning
action	[vx, vyaw]: forward velocity in m/s and yaw rate in rad/s. Positive vyaw is a right turn. In our experiments, cumulative (x, y, heading) waypoints relative to the current pose performed better than velocity targets; they can be reconstructed by integrating these 15 Hz commands in SE(2) from the start of each action chunk.
annotation.target_bbox_xyxy	Target box in 566 x 320 image pixels; left/top inclusive, right/bottom exclusive.
annotation.target_visible	Whether the target box is valid for that frame.
annotation.op_pixel_xy	Projected MAP point in image pixels.
annotation.op_visible	Whether the MAP point projects inside the view; this is not an occlusion test.

Parquet row t corresponds to video frame t. The video is AV1 in an MP4 container at 15 Hz. Ignore placeholder coordinates when their validity flag is false. Scene identities, source paths, world positions and headings, camera calibration, and geometric route records are not included. The bbox and MAP point coordinates are image pixels only.

Download

python -m pip install -U huggingface_hub
hf download Fysics-AI/MAP-Data --repo-type dataset \
  --include 'shards/map-data-*.zip' 'SHA256SUMS' 'LICENSE' \
  --local-dir ./MAP-Data
cd MAP-Data
sha256sum -c SHA256SUMS
unzip -q shards/map-data-00000.zip

Each shard can be verified and extracted independently. The UniWAM repository provides model training and MAP-Bench evaluation instructions.

Source and terms

This subset was rendered from HM3D scenes. The upstream Matterport terms and attribution requirements apply to derived media. See LICENSE. The other license label does not grant rights beyond those source terms.

Citation

Please cite the UniWAM paper and HM3D when using the data. Copy the UniWAM BibTeX entry below:

@article{xue2026uniwam,
  title   = {UniWAM Technical Report: Unified Mobile Manipulation via Mixed-Stream World-Action Modeling and Manipulation Anchor Pose Supervision},
  author  = {Wei Xue and Keliang Liu and Mingzhang Cui and Jinhua Xie and Jinjie Wei and Jianan Hou and Jingcheng Lu and Lintao Wang and Kaixiang Qiu and Yizhou Liu and Xinghai Ye and Jinghang Han and Mingcheng Li and Jie Gu and Shunli Wang and Lihua Zhang and Dingkang Yang},
  journal = {arXiv preprint arXiv:2609.39388},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.39388}
}