HUTACE-I5x64x64-B8x8-T65-0.83M

HUTACE is the Hybrid U-Net–Transformer Actor–Critic for Exploration. It selects local waypoints for energy-constrained coverage exploration from a partially observed 2.5D terrain.

The checkpoint has 833,795 parameters. It uses a 5×64×64 state, an 8×8 U-Net bottleneck, 64 spatial Transformer tokens plus one summary token (T65), two Transformer layers, and four attention heads.

• Interactive 3D demo: HUTACE 3D Explorer
• Frozen benchmark package: HUTACE Mapzen Benchmark

Input and output

Input is float32[5,64,64] in this order:

1. observed relative elevation, (elevation - start_elevation) / 200, clipped to [-3,3]
2. observed-cell mask
3. current-position one-hot map
4. visit count normalized as clip(count/3, 0, 1)
5. remaining-fuel ratio.

The model returns 4,096 waypoint logits and one critic value.

Quick use

from huggingface_hub import snapshot_download
import sys

repo = snapshot_download(
    "Moon-Young-Choi/HUTACE-I5x64x64-B8x8-T65-0.83M",
    revision="74e3bf02f3fb9636e2c418ac42f0251b2e25acfe",
)
sys.path.insert(0, repo)

from inference import load_model, select_action

model = load_model(repo)
action = select_action(state, action_mask, model=model)

state must be NumPy-compatible with shape [5,64,64]; action_mask must be Boolean-compatible with shape [4096]. For a compatible terrain NPZ containing elevation_m[64,64], use:

python run_exploration.py --terrain terrain.npz --seed 523 --output result.json

Frozen Mapzen test results

Method	Final coverage	Coverage–energy AUC	Coverage at 320	Energy to 50% (censored; lower is better)
HUTACE	0.7712	0.5213	0.5884	302.12
Plain U-Net PPO	0.6679	0.4511	0.5090	369.77
CNN PPO	0.5454	0.4062	0.4725	408.95
Nearest frontier	0.6114	0.3557	0.3636	494.37
Random local waypoint	0.4287	0.2684	0.2830	590.17
Expected gain / estimated energy	0.1405	0.1305	0.1402	612.55

HUTACE outperformed CNN PPO, Plain U-Net PPO, and all three listed partial-observation heuristics on these frozen-test means.

Terrain provenance

Training, validation, and test use Mapzen Terrain Tiles on AWS Open Data. Mapzen combines multiple regional source DEMs; it is not a single NASA or Copernicus dataset. Zoom-13 tiles are resampled onto a local-UTM 30 m simulator grid, which is not a universal native-resolution claim. Full source attribution is in docs/mapzen_attribution.md.