DeepFruits is a fruit detection system using deep neural networks, designed for autonomous agricultural robotic platforms. The system uses deep convolutional neural networks (DCNNs), specifically Faster R-CNN, to detect fruits in images from two modalities: RGB and Near-Infrared (NIR). The system combines multi-modal information through early and late fusion methods, achieving state-of-the-art results with an F1 score of 0.838 for sweet pepper detection, an improvement from 0.807. The system is efficient, requiring only bounding box annotations rather than pixel-level annotations, which significantly reduces annotation time. The model is retrained for seven fruits, with the entire process taking four hours per fruit. The system uses a pre-trained ImageNet model and adapts it for fruit detection. It combines RGB and NIR images to enhance detection accuracy, especially in challenging conditions like varying illumination and occlusions. The system is deployed on a laptop with an Intel i7 processor and a GeForce GTX 980M GPU, achieving an average processing time of 341 ms for JAI and 393 ms for Kinect 2 images. The system's performance is evaluated using precision-recall curves and F1 scores. Early and late fusion methods are compared, with late fusion outperforming early fusion. The system is tested on various fruits, including sweet peppers, rock melons, strawberries, apples, avocados, mangoes, and oranges, demonstrating robustness across different conditions and ripeness levels. The system is also tested in different environments, including commercial farms and grocery stores, showing its adaptability. The system's performance is compared with a CRF-based method, with Faster R-CNN showing superior results. The system's detection accuracy is influenced by the number of training images, with more images leading to better performance. The system is also tested in spatial-temporal independent conditions, showing its ability to generalize across different scenarios. The system's results demonstrate its effectiveness in detecting fruits in various conditions, with high accuracy and efficiency. The system is designed for real-time performance and is suitable for deployment in agricultural robotics. The system's open dataset and tutorial documentation encourage further research and development in the field of agricultural robotics.DeepFruits is a fruit detection system using deep neural networks, designed for autonomous agricultural robotic platforms. The system uses deep convolutional neural networks (DCNNs), specifically Faster R-CNN, to detect fruits in images from two modalities: RGB and Near-Infrared (NIR). The system combines multi-modal information through early and late fusion methods, achieving state-of-the-art results with an F1 score of 0.838 for sweet pepper detection, an improvement from 0.807. The system is efficient, requiring only bounding box annotations rather than pixel-level annotations, which significantly reduces annotation time. The model is retrained for seven fruits, with the entire process taking four hours per fruit. The system uses a pre-trained ImageNet model and adapts it for fruit detection. It combines RGB and NIR images to enhance detection accuracy, especially in challenging conditions like varying illumination and occlusions. The system is deployed on a laptop with an Intel i7 processor and a GeForce GTX 980M GPU, achieving an average processing time of 341 ms for JAI and 393 ms for Kinect 2 images. The system's performance is evaluated using precision-recall curves and F1 scores. Early and late fusion methods are compared, with late fusion outperforming early fusion. The system is tested on various fruits, including sweet peppers, rock melons, strawberries, apples, avocados, mangoes, and oranges, demonstrating robustness across different conditions and ripeness levels. The system is also tested in different environments, including commercial farms and grocery stores, showing its adaptability. The system's performance is compared with a CRF-based method, with Faster R-CNN showing superior results. The system's detection accuracy is influenced by the number of training images, with more images leading to better performance. The system is also tested in spatial-temporal independent conditions, showing its ability to generalize across different scenarios. The system's results demonstrate its effectiveness in detecting fruits in various conditions, with high accuracy and efficiency. The system is designed for real-time performance and is suitable for deployment in agricultural robotics. The system's open dataset and tutorial documentation encourage further research and development in the field of agricultural robotics.