MAP-Bench

Paper · Project · GitHub · Checkpoints

MAP-Bench evaluates instruction-conditioned navigation to a Manipulation Anchor Pose: the base position and heading from which a robot can begin the intended manipulation. Built in HM3D, it uses one forward RGB view and three distance levels: Short, Medium, and Long.

• 35,454 training trajectories with 7,184,098 RGB frames at 15 Hz, navigation actions, task instructions, target boxes, and MAP annotations.
• 15,000 evaluation episodes with scene references, start poses, target MAPs, and instructions.
• 75,000 evaluation rollouts and their results: 15,000 each for GR00T N1.7, π₀.₅, FastWAM, UniWAM-Base, and UniWAM-Scaled.

Level	Reference path length	Evaluation episodes
Short	< 1.5 m	2,110
Medium	1.5 m ≤ length < 3.5 m	10,014
Long	≥ 3.5 m	2,876
Total		15,000

The evaluation records are in evaluation/benchmark_15k.json, with matching IDs and distance levels in evaluation/trajectory_ids.csv. All five methods have one released rollout for each evaluation ID. Targets have five task paraphrases; historical rollouts retain the instruction used for that run. Use the saved task when replaying a released rollout.

Download

Run from your workspace directory to download the complete benchmark package:

python -m pip install -U huggingface_hub
hf download Fysics-AI/MAP-Bench --repo-type dataset --local-dir ./MAP-Bench

Training videos and evaluation rollouts are stored in compressed archives. Use the GitHub download helper with --training --traces to verify and extract them. Obtain HM3D scene assets separately for simulator evaluation.

Usage

See GitHub for benchmark setup and evaluation, navigation training, and results.

HM3D-derived data are subject to the source dataset terms; see LICENSE.md.

Citation

Please cite the UniWAM paper and HM3D when using MAP-Bench:

@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}
}