MSKit — Mini Simulation Kit

MSKit is a lightweight Python library for running terrain-based simulations backed by real-world elevation data from JAXA's AW3D30 30m Digital Elevation Model.

Tiles are streamed on-demand from the MegaBites-AI/AW3D30-DEM-Tiles dataset hosted on Hugging Face — no manual data download required.

Installation

pip install mskit

With optional visualisation support:

pip install mskit[viz]

Quick Start

from mskit import DEMLoader, RandomWalk, Projectile, WaterFlow, TerrainAgent

# Create a loader — tiles are fetched automatically as needed
loader = DEMLoader()

# ── Random Walk ──────────────────────────────────────────────────────────────
rw = RandomWalk(loader, start_lat=35.68, start_lon=139.69, step_m=300)
path = rw.run(steps=500)
print(f"Distance: {rw.total_distance_km():.2f} km")
print(f"Elevation gain: {rw.elevation_gain_m():.0f} m")

# ── Projectile ───────────────────────────────────────────────────────────────
proj = Projectile(loader, lat=36.0, lon=137.5,
                  azimuth_deg=45, elevation_deg=30, speed_ms=80)
traj = proj.run()
print(f"Range: {proj.range_km():.3f} km | Flight time: {proj.flight_time():.1f} s")

# ── Water Flow ───────────────────────────────────────────────────────────────
wf = WaterFlow(loader, patch_km=10)
flow = wf.run(lat=35.6, lon=137.5)
print(f"Flow length: {len(flow)} steps | Descent: {wf.total_descent_m():.0f} m")

# ── Terrain-Navigating AI Agent ──────────────────────────────────────────────
agent = TerrainAgent(loader,
    start_lat=35.60,  start_lon=139.70,
    target_lat=35.65, target_lon=139.75)
dataset = agent.generate_episode(max_steps=300, policy="mixed")
arr = agent.to_numpy()   # shape (N, 7): [step, lat, lon, elev, action, reward, done]
print(f"Episode: {len(dataset)} steps | Reached target: {agent.reached_target()}")

Simulations

Class	Description
`DEMLoader`	Multi-tile loader with LRU cache. Core data access layer.
`DEMTile`	Single 1°×1° elevation tile with spatial query methods.
`RandomWalk`	Slope-biased 2D random walk on real terrain.
`Projectile`	Ballistic trajectory simulation that terminates on terrain impact.
`WaterFlow`	D8 water runoff routing on real elevation grids.
`TerrainAgent`	RL-ready navigation agent producing training trajectories.

Dataset

Elevation tiles are served from:

MegaBites-AI/AW3D30-DEM-Tiles

Source: JAXA ALOS World 3D 30m (AW3D30 v3.2)
Coverage: Global
Resolution: 30 m/pixel, 1°×1° tiles (3600×3600 px)
Format: Compressed NumPy .npy (int16, metres)
Cache: Tiles are cached locally at ~/.cache/mskit/dem/

Uploading Tiles

To upload tiles to the dataset (MegaBites team only):

# Upload Japan tiles
python scripts/upload_tiles.py --region japan --token $HF_TOKEN

# Upload a custom bounding box
python scripts/upload_tiles.py --lat-range 30 45 --lon-range 130 145 --token $HF_TOKEN

# Dry run first
python scripts/upload_tiles.py --region japan --dry-run

License

MSKit code: MIT License

Produced by MegaBites AI