---

name: sensor-agent description: "Operate a trained TransFuser model through the CARLA sensor-agent interface, with checkpoint, sensor, tensor, control, and safety guidance that preserves the external CARLA runtime boundary." disable-model-invocation: true metadata: disco-role: operating license: MIT

TransFuser Sensor Agent

Use this sub-skill to prepare or diagnose the learned HybridAgent runtime for CARLA Leaderboard sensor-track evaluation. It covers the agent entry point, checkpoint-directory contract, backbone-dependent sensors, preprocessing, route-to-target transforms, CUDA placement, ensembling, PID control, and stuck/emergency behavior.

Route The Task

1. For entry-point, configuration-directory, checkpoint, ensemble, CUDA, or Leaderboard interface questions, read agent-runtime.md.

2. For sensor IDs, camera/LiDAR preprocessing, tensor shapes, coordinates, GPS route planning, or target-point construction, read sensor-and-tensor-contracts.md.

3. For action repetition, waypoint ensembling, NMS, PID conversion, stuck recovery, or emergency braking, read control-and-safety.md.

4. For missing CARLA/CUDA, bad args.txt, checkpoint-prefix mismatches, architecture drift, missing sensors, or route failures, read troubleshooting.md.

5. Before arranging a CARLA evaluation, run the bundled, side-effect-free preflight:

   python scripts/validate_agent_config.py /path/to/team-config
   

   Add --json for machine-readable output or --strict to make warnings fail. The validator parses JSON and checks checkpoint files statically; it does not import CARLA or PyTorch, deserialize weights, allocate CUDA memory, run inference, launch a server, or download anything.

Inputs And Outputs

• Input configuration is a directory containing args.txt and one or more .pth model state-dict files. The directory, not args.txt, is the Leaderboard TEAM_CONFIG value.
• The Python agent module must expose get_entry_point() returning exactly HybridAgent; HybridAgent implements the Leaderboard AutonomousAgent interface and returns carla.VehicleControl from run_step.
• The learned runtime is CUDA-only as implemented. CPU-only checks can validate files and schemas but cannot prove model inference.
• Route execution needs an externally supplied global plan, synchronized sensor frames, and a CARLA 0.9.10.1-compatible Python API/server stack.

Boundaries

This bundle deliberately does not launch CARLA, build evaluation commands, download checkpoints, run model inference, or modify checkpoints. Route full Longest6/local evaluation to the sibling carla-evaluation sub-skill, and route model training or checkpoint provenance to model-training.

The verified inspection environment had Python 3.7, CUDA, PyTorch 1.12.1+cu113, mmcv-full 1.6.0, mmdet 2.25.0, and timm 0.6.7; model, data, config, and train imports passed. The CARLA Python module was absent. Therefore preserve an explicit external dependency on CARLA 0.9.10.1, its Python API, a running server, and the compatible Leaderboard/scenario-runner stack. Do not interpret static validation as end-to-end simulator readiness.