sensor-agent
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.
- License 许可
- MIT
- 最近更新
- 2026-10-08
- Stars
- ★ 11
- Forks
- 3
01 前言 Front Matter
- name
- sensor-agent
- license
- MIT
- metadata
- { "disco-role": "operating" }
- 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
02 资源文件 4
- 参考文档references/agent-runtime.md7.8 KB
- 参考文档references/control-and-safety.md6.9 KB
- 参考文档references/sensor-and-tensor-contracts.md8.0 KB
- 参考文档references/troubleshooting.md9.7 KB
03 SKILL.md 正文
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
-
For entry-point, configuration-directory, checkpoint, ensemble, CUDA, or Leaderboard interface questions, read agent-runtime.md.
-
For sensor IDs, camera/LiDAR preprocessing, tensor shapes, coordinates, GPS route planning, or target-point construction, read sensor-and-tensor-contracts.md.
-
For action repetition, waypoint ensembling, NMS, PID conversion, stuck recovery, or emergency braking, read control-and-safety.md.
-
For missing CARLA/CUDA, bad
args.txt, checkpoint-prefix mismatches, architecture drift, missing sensors, or route failures, read troubleshooting.md. -
Before arranging a CARLA evaluation, run the bundled, side-effect-free preflight:
python scripts/validate_agent_config.py /path/to/team-configAdd
--jsonfor machine-readable output or--strictto 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.txtand one or more.pthmodel state-dict files. The directory, notargs.txt, is the LeaderboardTEAM_CONFIGvalue. - The Python agent module must expose
get_entry_point()returning exactlyHybridAgent;HybridAgentimplements the LeaderboardAutonomousAgentinterface and returnscarla.VehicleControlfromrun_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.
04 同仓库更多 Skills
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
python-api
Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP/mAP semantics, and source-style import diagnostics.
file-evaluation
Folder-based PASCAL VOC AP/mAP evaluation from ground-truth and detection text files.
object-detection-metrics
Guides legacy Object-Detection-Metrics PASCAL VOC AP/mAP evaluation from detection text folders and direct Python API classes.
nx-workspace-patterns
Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.