HammadNaseer/qcar2-ml-steering-weights

QCar2 ML Steering Weights

Weights and metadata for the QCar2 ROS 2 ML steering stack used in the rosbot_ws lane-keeping and LiDAR obstacle-avoidance project.

The primary model is an RF-DETR segmentation ONNX export used by rfdetr_onnx_lane_node to detect the lane mask and publish a steering target. The repository also includes fallback local weights used by the alternate YOLO/DL and ResNet lane-keeping modes in the workspace.

Files

File	Purpose
car_track_v3_lane.onnx	Primary RF-DETR ONNX segmentation model.
car_track_v3_lane.classes.txt	Class names in model output order.
class_names.json	Structured class-id mapping.
model_config.json	Input/output, runtime, and ROS usage metadata.
eval/onnx_runtime_smoke_test.json	ONNX Runtime smoke-test output and top scores.
eval/rfdetr_sample_predictions.json	Sample RF-DETR prediction from the project.
fallback/best.pt	Fallback YOLO/DL segmentation weight from the ROS package.
fallback/resnet18_road_following.pth	Fallback ResNet18 road-following weight.
checksums.sha256	SHA256 checksums for uploaded binary weights.
NOTES.md	Original local packaging and verification notes.

Class Names

Class ID	Name
0	background_class83422
1	lane2
2	traffic_light

Primary Model Details

• Architecture: RF-DETR segmentation medium
• Task: lane and traffic-light segmentation for QCar2 track perception
• Runtime: ONNX Runtime CPU or CUDA provider
• Input tensor: input, float32 NCHW [1, 3, 432, 432]
• Outputs:
  • dets: [1, 200, 4], normalized cx, cy, w, h
  • labels: [1, 200, 3], class logits
  • 4647: [1, 200, 108, 108], mask logits

Sample Scores

Sample project inference from eval/rfdetr_sample_predictions.json:

Class	Confidence	BBox XYXY
lane2	0.9720	[6, 99, 640, 480]

ONNX Runtime smoke-test top prediction from eval/onnx_runtime_smoke_test.json:

Class	Score	Normalized Box
lane2	0.8555	[0.5314, 0.5985, 0.9405, 0.7918]

These are smoke-test/sample scores, not a full benchmark.

Intended Use

This repository is intended for the QCar2 ROS 2 simulation/autonomy stack:

• RF-DETR ONNX lane target publishing
• ML lane keeping
• LiDAR-triggered obstacle avoidance
• one-way lane switching around obstacles

It is not a production autonomous-driving model and should not be used as a real vehicle safety system.

Download

hf download HammadNaseer/qcar2-ml-steering-weights \
  car_track_v3_lane.onnx car_track_v3_lane.classes.txt \
  --local-dir weights

ROS 2 Usage

Expected workspace paths:

weights/car_track_v3_lane.onnx
weights/car_track_v3_lane.classes.txt

The main script uses those files with:

scripts/run_autonomy.sh

or through:

LANE_KEEPER=rfdetr scripts/run_lane_keeping.sh

Checksums

c981e0652eb1c268f1fa28f7cc4d51e8df8b9064d78c7a3e3b84dfda0d762fbd  car_track_v3_lane.onnx
946e7817a5c809486fb5d467dfe1aca3746bf8684a9367894b3dbec9ebcdb7b8  car_track_v3_lane.classes.txt
5814d98ba5d313e624467d06bffc0b305e8586870fc431317f7a3f35451c060d  fallback/best.pt
774f4e7405ca050a7a8536e8eafa077fb18bbfc009364c3937d97d68866e0454  fallback/resnet18_road_following.pth

Limitations

• Trained/exported for this QCar2 simulation project and camera viewpoint.
• Sample scores are from local smoke tests, not a held-out validation set.
• Performance outside the lab track simulation is not guaranteed.
• License is marked other until the exact upstream dataset/export license is confirmed.