Running IntelX¶
Install IntelX and add the checkpoint first, as in Get Started.
One Episode¶
It explores until 99% of the free space is covered, drawing the run live the
way the paper's videos do: the predicted coverage-gain field in the world
frame, its ranked components, the junctions and their branches, the frontier
fallback, the goal and the decision in force. It writes the video,
final_trajectory.png and prediction_debug.log to results/run_intelx/.
Every knob lives in ExplorationConfig in src/intelx/config.py. The ones
that matter.
| Setting | Default | What It Does |
|---|---|---|
MODEL_CHECKPOINT |
weights_path("checkpoint_best.pth") |
the trained U-Net |
MODEL_TYPE |
resnet34 |
must match the checkpoint |
sim['MAP_PGM'] |
map_path("test/small_office3.pgm") |
the occupancy map to explore |
HEADLESS |
False |
True runs with no window, much faster |
VIEW_STYLE / VIEW_FPS |
'paper' / 10 |
the live view and how often it is drawn; 'debug' gives the older three-panel dashboard |
TURBO_MODE |
False |
True drops the artificial sleeps. FIXED_DT keeps the map identical, and the benchmark pins it on |
MAP_UPDATE_INTERVAL |
5 |
scan every N steps. Leave it at 5 for any reported run, it changes the coverage metric |
FIXED_DT |
0.05 |
pinned physics step, 20 Hz. None restores wall-clock timing |
LOCAL_ONLY_MODE |
False |
True disables junctions and the global planner, the Section V ablation |
WIN_M / LASER_SCAN_R |
20 / 10 |
prediction window and sensor range, in metres |
Any Single Method¶
python -m intelx.cli.run_method --list # or: intelx-method
python -m intelx.cli.run_method --method nearest --map ood/aws_maps/hospital.pgm
python -m intelx.cli.run_method --method ours --set BRANCH_ORDER=strongest
One run of any method on any map, drawn live in the same view. It is set up
exactly as one cell of the benchmark: the same pinned settings, the method's own
sensor, and the spawn the benchmark draws for that map and --seed, so what you
watch is what a row of results.csv measures. It writes
results/run_method/<map>_s<seed>_<method>.mp4, the final frame beside it and
the run's curves. --headless drops the window and --minutes caps the run;
Esc ends it early. --set KEY=VALUE overrides any setting, which is how the
ablations run: BRANCH_ORDER=strongest commits to the strongest branch first,
LOCAL_FIELD=geom_ig replaces the learned field with a geometric one, and
--method local_only drops the global planner.
Paths¶
Nothing in this tree stores an absolute path. Maps, weights and results resolve from the repository root, and each root takes an environment variable, which is how a run on a shared machine reaches a copy of the data without editing a single source file.