MAP-Bench

📄 Paper  |  🌐 Project  |  💻 Code  |  🤖 Models  |  📦 MAP-Data

MAP-Bench evaluates instruction-conditioned navigation to a Manipulation Anchor Pose (MAP), the base position and heading from which a robot can begin manipulation. The benchmark uses one forward RGB camera in HM3D and separates episodes by reference path length: Short (<1.5 m), **Medium** (1.5-3.5 m), and **Long** (>=3.5 m).

Release	Contents
Training split	35,454 trajectories, 7,184,098 RGB frames at 15 Hz, navigation actions, instructions, target boxes, and image-plane MAP annotations.
Evaluation split	15,000 episodes with instructions, scene references, starting poses, and goal MAP poses for closed-loop simulation.
Evaluation traces	75,000 rollouts: 15,000 each for GR00T N1.7, pi0.5, FastWAM, UniWAM-Base, and UniWAM-Scaled.

Benchmark results

Method	Position error (cm) ↓	Heading error (deg) ↓
GR00T N1.7	14.11	6.18
pi0.5	22.89	4.11
FastWAM	17.10	4.68
UniWAM-Base	12.51	3.24
UniWAM-Scaled	9.86	2.30

The values are equal-weight means of the three distance-level means over the same 15,000 evaluation episodes per method. Full level results are in results/summary.csv; per-episode results and trace checksums are in results/trajectory_metrics.csv. The project repository includes the paper's benchmark figures.

Download and evaluate

The UniWAM download helper retrieves relative workspace folders, verifies the archives, and extracts the training episodes and five sets of traces:

git clone https://github.com/Fysics-AI/UniWAM.git
cd UniWAM
python -m pip install -U huggingface_hub zstandard
python scripts/download_assets.py --benchmark --training --traces

The compressed dataset is about 18.3 GB. Obtain HM3D scene assets separately for closed-loop evaluation. See the repository's evaluation guide, navigation training guide, and model checkpoints.

Path	Purpose
training/archives/, training/index/	Compressed RGB training episodes and relative episode indices.
evaluation/benchmark_15k.json, evaluation/scene_contexts/	Evaluation episodes and HM3D scene context.
results/summary.csv, results/trajectory_metrics.csv	Aggregate and per-episode metrics.
archives/*.tar.zst	Five methods' raw evaluation traces, with archives/SHA256SUMS.

The evaluation files contain scene-relative world coordinates needed to render and score HM3D episodes. The training split contains the frame annotations and instructions described above. MAP-Bench uses only the navigation expert of the released UniWAM checkpoints; it does not evaluate manipulation. For a separate, larger HM3D-derived video and frame annotation subset, see MAP-Data.

Source and terms

HM3D-derived content remains subject to the upstream HM3D and Matterport terms. Users obtain HM3D scene assets from their publisher. See LICENSE.md, and cite the UniWAM paper and HM3D when using the benchmark.