RESEARCH · RESEARCH · #848
PAANI: On-device visual evidence fusion and explainable guidance for river-robot simulation (arXiv:2609.22353v1)
PAANI is an on-device perception-to-guidance architecture for river-monitoring robots that fuses timestamp-aligned outputs from a project-trained YOLO11n detector and a MobileNetV3 Small semantic segmenter on an Arduino UNO Q, exposing the evidence and policy reasons behind each advisory. Models were trained on 10,000 WaterScenes detection images and 1,127 MaSTr1325 segmentation images (validation: 198 seg images); the FP32 ONNX exports total 14.817 MB, detector mAP@0.5 ≈ 0.7388/0.7367, segmentation mIoU = 0.9750, and a five-minute UNO Q run produced median/95th-percentile pipeline latencies of 467.8 ms / 580.3 ms at 0.5 Hz, while evaluations also identified black-input misclassification and a sampling-rate mismatch that limited apparent-motion evidence collection.
KEY POINTS
- PAANI is an on-device perception-to-guidance architecture for river-monitoring robots that fuses timestamp-aligned outputs from a project-trained YOLO11n detector and a MobileNetV3 Small semantic segmenter on an Arduino UNO Q, exposing the evidence and policy reasons behind each advisory.
- Models were trained on 10,000 WaterScenes detection images and 1,127 MaSTr1325 segmentation images (validation: 198 seg images); the FP32 ONNX exports total 14.817 MB, detector mAP@0.5 ≈ 0.7388/0.7367, segmentation mIoU = 0.9750, and a five-minute UNO Q run produced median/95th-percentile pipeline latencies of 467.8 ms / 580.3 ms at 0.5 Hz, while evaluations also identified black-input misclassification and a sampling-rate mismatch that limited apparent-motion evidence collection.
- The paper demonstrates an inspectable, low-footprint on-device AI pipeline for river robotics and quantifies trade-offs between model accuracy and on-board execution while highlighting remaining validation and sensor-sampling issues.
WHY IT MATTERS
The paper demonstrates an inspectable, low-footprint on-device AI pipeline for river robotics and quantifies trade-offs between model accuracy and on-board execution while highlighting remaining validation and sensor-sampling issues.