---

name: data-and-routes description: "Guides TransFuser CARLA training route and scenario preparation, dataset-layout validation, privileged autopilot collection, safe command construction, and route visualization without launching CARLA or downloading data." disable-model-invocation: true metadata: disco-role: operating license: MIT

TransFuser Data And Routes

Use this sub-skill when a Researcher needs to understand, validate, or prepare TransFuser training routes, scenario annotations, generated samples, or the privileged data-generation agent. Start with the relevant bundled reference; use the scripts for deterministic local checks and command construction.

Route by need

1. Existing files or schema: read dataset-layout.md for sample folders and routes-and-scenarios.md for XML/JSON validation and route semantics. Run: python scripts/validate_route_files.py --help and python scripts/validate_dataset_layout.py --help.
2. Generating or visualizing files: read data-generation.md, then build a plan with python scripts/build_route_tool_command.py --help. The builder only prints a command/environment plan; it never starts a server, imports CARLA, writes routes, or uses the network.
3. A failed or incomplete run: read troubleshooting.md, validate inputs before retrying, and preserve the distinction between a safe local check and a CARLA-backed result.

Required operating sequence

• Confirm the town set (Town01–Town07, Town10HD), scenario set (Scenario1, 3, 4, 7, 8, 9, 10), output roots, and whether the operation is route generation, scenario generation, visualization, or collection.
• Validate paths and schemas first. Route XML and scenario JSON can be parsed on CPU without CARLA; route/scenario compatibility is a structural signal, not proof that a CARLA map can interpolate the route.
• Generate scenarios before routes. Start the matching CARLA 0.9.10.1 server and then run the scenario launcher, followed by the route launcher. Never run the bundled command builder as if it were a launcher.
• Keep PYTHONPATH ordered for the CARLA PythonAPI/egg, ScenarioRunner, and leaderboard. Use absolute paths at execution time, but do not place a local checkout path into a reusable skill file.
• For collection, use the privileged DataAgent/AutoPilot only after the route/scenario pair is validated, the server is reachable, and the output directory has enough space. A generated command or valid fixture is never a claim that data generation passed.
• Stop before network downloads, simulator launch, Docker, or the 210 GB data acquisition boundary unless the Researcher explicitly owns those external actions and verifies them separately.

Invariants and handoffs

• data-and-routes owns route/scenario generation, route visualization, collection layout, and safe preflight. Link to model-training for how samples are consumed and to carla-evaluation for Longest6/evaluator runs.
• The generated dataset must contain synchronized frame indices across the required modalities and enough future label_raw frames for the configured sequence/prediction window.
• Scenario files describe trigger transforms and event configurations; route files describe map/town and waypoint paths. A matching filename alone does not prove that an event trigger lies on a route.
• The repository checkout is source evidence only. These bundled references and scripts must remain usable when the original checkout is absent.

Verification status

The route/scenario schemas, selected repository fixtures, and self-contained script fixtures are CPU-checkable. CARLA 0.9.10.1 and a running CarlaUE4.sh server were absent during construction, so no scenario generation, route interpolation, visualization against a map, or dataset collection is claimed as passed. Treat this as an explicit external-runtime block.