DanielDDDS/topological-traps-dataset

Topological Traps Dataset

TAU Algorithmic Robotics - Fall 2025/2026 - Daniel Simanovsky

Pre-processed occupancy grids and Oracle viability labels derived from the HouseExpo residential floor-plan dataset, used to train DanielDDDs/topological-traps.

Contents

Path	Description	Size
data/processed/	1,001 binary occupancy grids (512x512 px, .npy)	~262 MB
data/manifest.csv	Train/val/test split (700/150/150), seed 42	<1 MB
data/labels/robot_20x15/	Oracle viability labels, small robot	~1 GB
data/labels/robot_30x20/	Oracle viability labels, default robot	~1 GB
data/labels/robot_40x25/	Oracle viability labels, large robot	~1 GB
data/labels/robot_25x18/	Oracle viability labels, unseen test robot	~1 GB

Label format

Each label file is a (4, 512, 512) uint8 NumPy array. Channel order: [North, South, East, West]. 1 = viable (robot can escape), 0 = directional trap.

Robot sizes

Size (LxW px)	Diagonal (px)	Split
20x15	25	Train
30x20	36	Train
40x25	47	Train
25x18	31	Test only (unseen)

Oracle algorithm

1. Rotation-safe mask: erode free space with circular kernel of diameter sqrt(L^2+W^2)
2. Translation-safe masks: erode with oriented rectangular footprint per direction
3. Reverse BFS flood-fill: seed rotation+translation-safe pixels, propagate backwards

Loading example

import numpy as np

occ = np.load("data/processed/0041a20dcdfd5e0d1ca0752365a70634.npy")
# shape (512, 512), uint8, 1=free 0=wall

label = np.load("data/labels/robot_30x20/0041a20dcdfd5e0d1ca0752365a70634.npy")
# shape (4, 512, 512), uint8
north_viable = label[0]

Source

1,000 maps from HouseExpo (Li et al., 2019, arXiv:1903.09845), rasterised to 512x512.

Code

github.com/danielsddd/topological-traps-project