create-scenic-scenario

Writes a new Scenic scenario (.scenic) from a natural-language description — "a car cuts in from the right lane", "a pedestrian crosses in front of the ego", "negotiate an unsignalized four-way" — by

SKILL.md

GitHub

Create a Scenic scenario from a description

Paths. scripts/… and references/… below are relative to the directory holding this SKILL.md. Your working directory is the user's project, not that directory, so prefix them with its absolute path or the command is not found.

Scenic describes a distribution over scenes, not one scene. So a good scenario is not a script of positions — it is a set of constraints loose enough to sample and tight enough to be the situation asked for.

Translating a request has three parts, and only the first is a judgement call:

  1. request → spec (yours): pick the map, the placement, the actors and their relations.
  2. spec → file (scripts/scaffold_scenic.py): emits boilerplate and placement idioms lifted from scenarios that run on this build.
  3. validate (same script): compile, then sample. No server needed.

Do not hand-write the boilerplate. Three things bite every time and the generator already handles them:

  • assert len(x) > 0 is illegal on a random value. Anything derived from a Uniform(...) raises RandomControlFlowError. Only assert on lists built deterministically from network.
  • Pedestrians and props need with regionContainedIn None when placed on the road. Containment is re-checked every sample, so without it generation never converges — it looks like an unsatisfiable require.
  • param map and param carla_map must name the same town. map builds the road network; carla_map tells CARLA what to load. Disagree and the scene is sampled on one map and simulated on another.

Instructions

Progress:
- [ ] Step 1: Check prerequisites (bash scripts/check_env.sh)
- [ ] Step 2: Pick a map that can express the situation
- [ ] Step 3: Pick blueprints that exist in THIS build
- [ ] Step 4: Write the spec, generate, and let it validate
- [ ] Step 5: Loop on the verdict until it samples; then simulate

Step 2: The map has to support the situation

A scenario selects road features. If the map has none of them, no wording will save it — the generated filter matches nothing and compilation fails with tried to make discrete distribution over empty domain.

python3 ../run-scenic-scenario/scripts/list_scenic.py --check-maps

Read the 4way=N(uns M) / 3way=N(uns M) columns. Junction-negotiation scenarios need uns > 0; dense downtown maps are usually fully signalized.

Step 3: Blueprints that exist here

python3 scripts/blueprint_table.py                      # inventory + gaps
python3 scripts/blueprint_table.py --category bicycle   # ids Scenic offers
python3 scripts/blueprint_table.py --list vehicles      # what the build has
python3 scripts/blueprint_table.py --check vehicle.lincoln.mkz vehicle.lincoln.mkz_2017

Two lists, and they disagree. blueprint_table.py reads the build's own Content/Carla/Config/*Parameters.json — the definitive inventory — and diffs it against Scenic's table for the installed client version.

  • A category Scenic reports empty makes new Bicycle fail at sample time even when the build has bicycles. Name the id instead: with blueprint "vehicle.bh.crossbike".
  • An id in the build but absent from Scenic's table is still spawnable — reference it explicitly.
  • An id in neither fails at spawn time, long after sampling passed. This is how the upstream scenarios still carrying vehicle.lincoln.mkz_2017 behave.

Step 4: Spec, generate, validate

python3 scripts/scaffold_scenic.py --example > spec.json   # start from this shape
# edit spec.json
python3 scripts/scaffold_scenic.py --spec spec.json --out my.scenic

Generating with --out validates automatically. To re-validate after a hand edit:

python3 scripts/scaffold_scenic.py --validate my.scenic

Spec keys:

Key Meaning
map town name; needs a matching .xodr in the assets (required)
model srunner (default) or scenic — see [[run-scenic-scenario]]
placement lane for road situations, intersection for junction ones
arms, signalized intersection filter; signalized: false for negotiation
ego {blueprint, speed, maneuver}; maneuver is straight/left/right
actors[] {name, type, blueprint, relation, distance, speed, maneuver, offset, threshold}
requires[] extra require lines, verbatim
terminate termination condition; defaults to leaving the spawn area

relation decides the geometry, and each maps to a proven idiom:

relation Geometry
ahead / behind following roadDirection from ego for ±Range(lo, hi)
right_lane / left_lane selects lane sections with that neighbour, places there
conflicting a conflicting maneuver through the same intersection
crossing anchored crossing: ego starts back up the lane, actor crosses it

Step 5: Loop on the verdict

Verdict Meaning Fix
PASS sampled in N iterations done — simulate it [[run-scenic-scenario]]
COMPILE-FAIL ... empty domain map lacks the feature change map, or relax arms/signalized
COMPILE-FAIL RandomControlFlowError a comparison on a random value hand-edited assert — remove it
COMPILE-FAIL ... no 'X' blueprints recorded Scenic category empty name a concrete blueprint
SAMPLE-FAIL RejectionException requirements jointly unsatisfiable loosen the tightest require, widen distance

Simulating is a separate, explicitly-requested step — generating a validated scenario is this skill's deliverable. Run it with [[run-scenic-scenario]].

Examples

Example 1: "a car cuts in front of me from the right lane"

placement: lane, one actor with relation: right_lane. The generator selects lane sections that have a right neighbour, because only sections know their neighbours — network.lanes does not. Samples in single-digit iterations on Town05.

Example 2: "a pedestrian steps into the road ahead of the ego"

relation: crossing, type: Pedestrian. Uses CrossingBehavior(ego, speed, threshold) so the walker waits until the ego is threshold metres away, and adds regionContainedIn None. relation: ahead with a Pedestrian would sample forever instead.

Example 3: "negotiate an unsignalized four-way with a car on a conflicting arm"

placement: intersection, arms: 4, signalized: false, one actor with relation: conflicting. Check --check-maps first: on a map whose only four-way is signalized this cannot compile, which is exactly why the shipped carlaChallenge10 fails on Town10HD_Opt.

Troubleshooting

Problem: RandomControlFlowError: random values cannot be compared Cause: a comparison — usually assert len(...) > 0 — on something derived from a Uniform(...). Scenic evaluates the scenario symbolically, so control flow cannot depend on a sample. Solution: assert only on lists built by a plain loop over network.*. The generator never emits the illegal form.

Problem: sampling never converges on a scenario with a pedestrian or a prop Cause: the default containment region. The object is being placed on the road but must lie in the walkable region, so every sample is rejected. Solution: with regionContainedIn None.

Problem: no <Town>.xodr found Cause: the generator resolves the OpenDrive from SCENARIO_RUNNER_ROOT/srunner/scenic/assets and a Scenic checkout's assets/maps/CARLA. Solution: set those roots, or pass "xodr" in the spec explicitly.

Problem: the scenario samples but nothing spawns when simulated Cause: a blueprint id that does not exist in this build. Ids are resolved at spawn time, so sampling cannot catch it. Solution: scripts/blueprint_table.py --check <id> ... on every id in the file.

Problem: a hand-written terminate when never fires Cause: it references an object that has moved, or a distance that never grows. Solution: keep the generated form — distance from the spawn point, which always grows once the ego drives — and always bound the run with --time anyway.

Outputs

A .scenic file that compiles and produces at least one sampled scene, with the blueprints of that scene printed so they can be checked against the build. The file is ready for scenic FILE --simulate --2d via [[run-scenic-scenario]].

The language constructs this model actually supports — object classes, behaviors, spatial operators, requirement forms — are in references/language.md.