IntelX Docs

How much new space would I gain by standing there?¶

IntelX answers that question for every nearby goal with one lightweight U-Net, and explores unseen indoor spaces on the junctions it reads from the answer.

These pages cover installing IntelX, running it and every baseline in one simulator, reproducing the benchmark, and training the coverage-gain regressor.

Get Started Run the Benchmark Project Page

IntelX overview: a robot in an unseen 2D environment builds its map from a 220-degree laser scan; a robot-centred 20 m window of that map goes through a U-Net that returns a coverage-gain heatmap; the heatmap's angular components are candidate directions, and multi-peaked structure marks a junction with branches; the local and global planner commits to those branches, falls back to frontier centroids when it is stuck, and explores the whole floor along a short path. — From the robot's scan and local window to the predicted coverage gain, its candidate directions and junctions, and the explored floor.

Try It Now¶

git clone https://github.com/IRVLUTD/IntelX.git
cd IntelX
python3.11 -m venv .venv && . .venv/bin/activate
pip install -r requirements.txt && pip install -e .
python -m intelx.cli.run_sweep_all --selftest      # the benchmark's self-test, no GPU

The self-test runs the benchmark's own helpers, from spawn sampling and the coverage metrics to task splitting and CSV writing, with no simulator, checkpoint or GPU, so it is the fast way to check an installation. To watch IntelX explore, add the checkpoint and follow Get Started.

What the Paper Found¶

>20%Shorter Path to 99% Coverage Than Every Baseline That Finishes the Larger Maps

8.1 msPer 20 m Coverage-Gain Field, About 120 FPS

10⁶Training Samples, Each Labelled by a Simulation Oracle

8Baselines in One Simulator, Run From the Same Spawns

From the paper: the larger held-out maps are Medium Office 6 and Large Office 7 (Table 3), and the speed is Table 2's.

All nine methods exploring KTH Large Office 7 from the same spawn, replayed side by side and synchronized by distance travelled, with their coverage curves below. IntelX reaches 99% coverage at 1,687 m; IG-Hector, UPEN, MapEx and Cost-Utility reach it later, and Random Frontier, Nearest Frontier, Geometric IG and Frontier (AutoX) stop short. — All nine methods on KTH Large Office 7 from the same spawn, synchronized by distance travelled. The project page replays the other maps too.

How It Fits Together¶

Every method runs in one 2D simulator that casts its laser rays in CUDA kernels. IntelX's regressor, a U-Net with a ResNet-34 encoder, predicts over a robot-centred 20 m window how much new area the sensor would gain from each nearby goal. The planner reads angular components of that field as candidate directions, reads multi-peaked structure as a junction, and commits to branches weakest first. The global planner keeps each junction's branches and, when the window offers nothing, heads for the nearest active branch or frontier. The benchmark draws one spawn per map and seed and runs every method from it.

Key	Method	Notes
`ours`	IntelX	the coverage-gain field, the local and the global planner
`local_only`	IntelX, Local Only	the ablation without the global planner
`random` `nearest` `cost_utility` `geom_ig`	Random Frontier, Nearest Frontier, Cost-Utility, Geometric IG	geometry only
`frontier`	Frontier (AutoX)	our previous work, a dynamic-window frontier explorer
`mapex` `upen` `hectoraug`	MapEx, UPEN, IG-Hector	need MapEx and its service

Where to Go Next¶

01Get StartedInstall, add the checkpoint and the maps, and run IntelX. 02Running IntelXOne episode, any single method, the settings and the paths. 03The BenchmarkThe paired protocol, every baseline and the MapEx service. 04Training the RegressorCollect coverage data, train the U-Net and evaluate it. 05Code MapThe layout, every module and the tests. 06Citation and LicensesThe citation, coming with the arXiv paper, and the licenses of what IntelX builds on.