moeen14/argus-omnimotus-vision-models

Argus Omnimotus Vision Models

Vision models for Argus Omnimotus, an open-source agricultural robotics platform. The release supports crop and platform identification, overhead task verification, rail segmentation, and optional prompt-based segmentation.

The robot software, deployment tools, hardware files, and operating instructions are available in the Argus Omnimotus GitHub repository.

Models

File	Task	Classes
crop_detector.onnx	Bottom-camera object detection	gCrop, platform, ycrop, servo
highcam_platform.onnx	Overhead object detection and task verification	ball, gcrop, platform, ycrop
highcam_rail.onnx	Overhead rail segmentation and BEV navigation	rail
lanesegyv8.onnx	Earlier rail-segmentation model retained for reproduction	rail
resnet18_image_encoder.onnx	Optional NanoSAM image encoder	prompt-based segmentation
mobile_sam_mask_decoder.onnx	Optional NanoSAM mask decoder	prompt-based segmentation

The six files above are the canonical models consumed by the robot repository. Their sizes and SHA-256 hashes are recorded in manifest.json.

The edge/ directory contains the YOLO model variants evaluated in the paper:

Directory	Contents
edge/onnx/	Portable FP32 reference exports
edge/hailo8/	INT8 HEF files compiled for Hailo-8
edge/ncnn/	FP32, FP16, INT8, and mixed-INT8 NCNN exports for Raspberry Pi 5

TensorRT engines are intentionally generated on the target Jetson because compatibility depends on the installed TensorRT, CUDA, and GPU environment. The GitHub repository includes the build and evaluation commands for Raspberry Pi 5, Hailo-8, and Jetson Orin Nano.

Input and preprocessing

The YOLO models use 1 x 3 x 640 x 640 RGB input normalized to [0, 1]. Letterboxing, output decoding, non-maximum suppression, and task-specific thresholds are implemented in the linked deployment code. NanoSAM uses its upstream input and prompt conventions.

Download

hf download moeen14/argus-omnimotus-vision-models --local-dir models

For the robot runtime, download and verify only the six canonical ONNX files with the repository utility:

python tools/models.py download --repo-id moeen14/argus-omnimotus-vision-models
python tools/models.py verify
python tools/models.py build

Intended use

These models are intended for research and education with the Argus Omnimotus testbed. Performance can change with camera placement, crop geometry, lighting, soil appearance, and accelerator toolchain versions. Validate confidence thresholds and safety behavior on the target platform before autonomous operation.

Authors and citation

The Argus Omnimotus project was developed by Moeen Ul Islam, Bishal Adhikari, J. Alex Thomasson, and Dong Chen. Please cite the project using the CITATION.cff file in the GitHub repository.

License and attribution

The YOLO-derived models are released under the GNU Affero General Public License v3.0. The NanoSAM encoder and decoder retain their Apache-2.0 upstream terms; see THIRD_PARTY.md. Users of these files must comply with the license applicable to each model and credit the Argus Omnimotus project in resulting research and derivative work.