PAANI
An on-device visual evidence fusion architecture designed for navigation decision-making in river-monitoring robots. It integrates object detection and semantic segmentation to provide interpretable navigation recommendations, where each instruction is accompanied by supporting evidence and policy reasoning.
Timeline (last 90 days)
PAANI is an on-device architecture spanning perception to navigation, deployed on resource-constrained hardware such as the Arduino UNO Q, combining a self-trained YOLO11n detector and a custom MobileNetV3 Small semantic segmenter.
PAANI's YOLO11n detector handles object detection, and the MobileNetV3 Small semantic segmenter handles water surface region delineation, with the outputs of both integrated through a timestamped-aligned evidential fusion mechanism.
Every final navigation recommendation from PAANI exposes the supporting evidence and the reasoning behind the strategy, achieving inspectability.
PAANI connects the local AI pipeline to an independent Gazebo simulated vessel, localization, and control testbed via a ROS 2 interface, forming a closed-loop simulation validation environment.
PAANI's detection task used 10,000 WaterScenes images for four-class object detection training, and the segmentation task used 1,127 MaSTr1325 images (including 198 segmentation validation images).
The FP32 ONNX model ultimately selected by PAANI occupies only 14.817 MB.
The PAANI detector checkpoint achieved a test mAP of 0.7388 at 0.5 IoU, and the rectangular ONNX export version achieved a validation mAP of 0.7367.
The PAANI segmentation ONNX achieved a validation mIoU of 0.9750.
A five-minute recording conducted on the UNO Q at a 0.5 Hz sampling rate showed a pipeline latency median of 467.8 milliseconds and a 95th percentile of 580.3 milliseconds.
Two explicit issues were identified in PAANI's evaluation: misclassification of black inputs (pure black or extremely dark frames leading to incorrect judgments), and a sampling rate mismatch causing the apparent motion estimator to be unable to collect sufficient evidence.
2 more timeline events
All Facts (12)
PAANI is an on-device architecture spanning perception to navigation, deployed on resource-constrained hardware such as the Arduino UNO Q, combining a self-trained YOLO11n detector and a custom MobileNetV3 Small semantic segmenter.
50%UnverifiedPAANI's YOLO11n detector handles object detection, and the MobileNetV3 Small semantic segmenter handles water surface region delineation, with the outputs of both integrated through a timestamped-aligned evidential fusion mechanism.
50%UnverifiedEvery final navigation recommendation from PAANI exposes the supporting evidence and the reasoning behind the strategy, achieving inspectability.
50%UnverifiedPAANI connects the local AI pipeline to an independent Gazebo simulated vessel, localization, and control testbed via a ROS 2 interface, forming a closed-loop simulation validation environment.
50%UnverifiedPAANI's detection task used 10,000 WaterScenes images for four-class object detection training, and the segmentation task used 1,127 MaSTr1325 images (including 198 segmentation validation images).
50%UnverifiedThe FP32 ONNX model ultimately selected by PAANI occupies only 14.817 MB.
50%UnverifiedThe PAANI detector checkpoint achieved a test mAP of 0.7388 at 0.5 IoU, and the rectangular ONNX export version achieved a validation mAP of 0.7367.
50%UnverifiedThe PAANI segmentation ONNX achieved a validation mIoU of 0.9750.
50%UnverifiedA five-minute recording conducted on the UNO Q at a 0.5 Hz sampling rate showed a pipeline latency median of 467.8 milliseconds and a 95th percentile of 580.3 milliseconds.
50%UnverifiedTwo explicit issues were identified in PAANI's evaluation: misclassification of black inputs (pure black or extremely dark frames leading to incorrect judgments), and a sampling rate mismatch causing the apparent motion estimator to be unable to collect sufficient evidence.
50%UnverifiedPAANI currently validates model accuracy and on-board execution capability; collision avoidance performance in real waterways still requires further validation.
50%UnverifiedPAANI introduces bounded tracking to provide continuous target recognition, avoiding judgment fluctuations caused by single-frame missed detections, and employs an explicit channel strategy that synthesizes water surface labels, detection results, urgency, and segmentation uncertainty to produce navigation recommendations.
50%