Researchers propose PAANI, an on-device visual evidence fusion architecture for river robot guidance

Researchers from Savio Cardoz and Santhiya Rajan introduce PAANI, an on-device perception-to-guidance architecture designed for mobile river monitoring robots operating under resource constraints. The system targets a gap where geographic waypoints alone cannot capture dynamic obstacles or water boundaries, and where conventional object labels or steering commands fail to explain which evidence supports a decision or when that evidence is unreliable.

PAANI integrates a project-trained YOLO11n detector and a custom MobileNetV3 Small semantic segmenter with timestamp-aligned evidence fusion on an Arduino UNO Q platform. Bounded tracking supplies object persistence, while an explicit corridor policy combines surface labels, accepted detections, urgency, and mask uncertainty. Each final advisory exposes its contributing evidence and policy reasons, aiming to make decisions inspectable on-device.

Training data includes 10,000 WaterScenes images for four-class detection and 1,127 MaSTr1325 images for segmentation, with 198 images reserved for segmentation validation. The selected FP32 ONNX models occupy 14.817 MB of storage. The detector checkpoint reports a test mean average precision at 0.5 IoU of 0.7388, while a separately evaluated rectangular ONNX export achieves a validation mAP at 0.5 IoU of 0.7367. The segmentation model achieves an ONNX validation mean intersection-over-union of 0.9750.

On-hardware evaluation on a five-minute Arduino UNO Q recording, configured at a 0.5 Hz cadence, reports median and 95th percentile pipeline latencies of 467.8 ms and 580.3 ms respectively. The authors also document limitations: black input misclassification and a sampling rate mismatch that prevents the diagnostic apparent motion estimator from collecting sufficient evidence. Despite these issues, the results are presented as supporting an inspectable and reusable edge robotics foundation, while distinguishing model accuracy from validated on-water collision avoidance performance.

The paper includes system and AI architecture diagrams, model-training results, qualitative evaluations, and deployment measurements. The authors state that project code, trained models, ONNX artifacts, logs, and reproducibility documentation are available at a provided URL.