Skip to content

SCENES.md

Latest commit

 

History

History
43 lines (35 loc) · 3.54 KB

SCENES.md

File metadata and controls

43 lines (35 loc) · 3.54 KB
Raw

Cesium viewer — bundled demo scenes

These public/*.json files are pre-rendered, self-contained scenes (each embeds the SF terrain reference, roads, buildings, and the full per-step car/pedestrian trajectory). They are committed so teammates can view the demos without re-rendering — just start the viewer and pick a scene from the dropdown.

View them

cd smoothride/demo/cesium
python3 -m http.server 8137
# open http://127.0.0.1:8137  → use the "iter" dropdown (top-left HUD) to switch scenes

Needs a Cesium ion token in the git-ignored config.js (copy config.example.js and paste a free token). Without a token the viewer still works (flat ellipsoid + GeoJSON buildings instead of photoreal terrain).

Reading the viewer: cars are cones — red = crashed, green = arrived, blue = en-route (brighter = faster). Pedestrians are amber cylinders. Press play on the Cesium timeline; zoom into an intersection to watch cars slow for crossing pedestrians and queue without colliding. The HUD shows live cars / trips / crashed counts.

The dropdown (manifest.json)

manifest.json lists the scenes and their labels; the viewer builds the dropdown from it and loads the last entry by default. Edit/extend it to add scenes ({"iter", "file", "label"}).

What each scene is

Dropdown label File Model / checkpoint Region Cars / Peds Shows
iter 0 (baseline) scene_it00000.json trained_peds @ iter 0 (untrained) downtown 96 / 300 Starting point — cars barely move, many crashes
iter 50 … 250 scene_it000{50,100,150,200,250}.json trained_peds @ that iteration downtown 96 / 300 Training progression — scrub to watch the policy learn
iter 299 scene_it00299.json trained_peds (final) downtown 96 / 300 Fully-trained v1 (dense 300-ped downtown)
CHAMPION v4loo — held-out Mission scene_champion_mission.json trained_v4loo (v2 generalization champion, trained on downtown+nopa+chinatown — never saw Mission) mission 96 / 10 Cross-region generalization — ~1–2% crashes on an unseen neighborhood

Notes:

  • The iter 0…299 series is the v1 model (single-cost, dense pedestrians) — kept because it's the clearest "watch it learn" progression for the dropdown.
  • The champion Mission scene is the v2 leave-one-out model — the headline generalization result (see top-level README.md and docs/internal/HANDOFF-overnight.md).
  • Scene files are large (~3.6 MB each) and public/ is otherwise git-ignored; these specific demo scenes were force-added. New ad-hoc renders won't be auto-committed.

Re-render / add scenes

Checkpoints live in the Modal volume smoothride-nav-ckpts (and runs/ locally). To render a checkpoint to a scene for a region:

# example: champion v4loo on Mission
modal volume get smoothride-nav-ckpts trained_v4loo.msgpack runs/trained_v4loo.msgpack
mkdir -p runs_demo && cp runs/trained_v4loo.msgpack runs_demo/trained_demo_it00000.msgpack
python3 scripts/export_snapshots.py --region mission --tag _demo \
  --agents 96 --n-peds 10 --steps 250 --elevation synthetic \
  --ckpt-dir runs_demo --out-dir smoothride/demo/cesium/public
# then add an entry to public/manifest.json pointing at the new scene_*.json

scripts/export_snapshots.py renders a whole series of versioned checkpoints + writes a manifest.json; smoothride/demo/export_cesium.py renders a single scene. Available regions: downtown, mission, nopa, chinatown_fidi (smoothride/data/map_loader.py::SF_REGIONS).