2410.21153 Synthetica: Large Scale Synthetic Data for Robot Perception

synthetic data for agricultural robotics

The camera also was set on a track and animated to move along it, changing angles and perspectives in each frame. Finally, the Principled BSDF node was used to increase the roughness of the leaf’s surface, further improving its realism. A shader nodes setup was created that allowed for variation in leaf color between lighter and darker green. To further enhance the model’s variability, the random node was incorporated into the node network, enabling the randomization of parameters such as leaf size and orientation, stem curvature, and strawberry color. Flowers were created using Bézier curves for the petal and calyx structures, which were further adjusted using rotation, scaling, and bending modifiers to ensure a realistic appearance. Similarly, fruits were modeled as a series of stages from unripe to ripe, using a base geometric shape modified by noise and texture gradients to mimic the natural appearance of strawberries at different 2026 responsible AI stages of growth.

DDC has been successfully applied in various fields, including medical imaging, where models trained on one set of MRI scans were effectively adapted to new datasets from different MRI machines, significantly improving performance . One of the most effective domain adaptation techniques is Deep Domain Confusion (DDC) , which addresses domain shift by integrating a domain confusion loss into the training process. Domain adaptation techniques offer a solution to this issue by aligning the distributions of source and target domains, enhancing model performance without requiring extensive retraining on new data . While SwinUNet is commonly used in medical imaging, it has recently AI Index data and research been applied to agricultural tasks such as fruit ripeness detection and crop segmentation .

The use of DDC techniques helped bridge this gap to some extent, yet the model still benefited greatly from the increased diversity and quantity of synthetic training data. U-Net-based architectures, while effective in biomedical applications, often exhibit reduced accuracy in 2026 AI economy agricultural occlusions . Meanwhile, YOLO-based models such as LS-YOLOv8s are well-suited for real-time object detection but lack fine-grained pixel-wise segmentation capabilities. To contextualize these results, this approach can be compared with existing state-of-the-art segmentation models commonly used in agricultural applications. The results of this study highlight the effectiveness of the SwinUNet architecture combined with deep domain confusion (DDC) techniques for segmenting ripe and unripe strawberries from synthetic training data. The trained model was tested on real images from the StrawDI dataset , captured using Samsung Galaxy S7 Edge smartphone.

synthetic data for agricultural robotics

Fig 4. Example of grayscaled real-world image (left) alongside corresponding segmentation mask (right).

Leaves were modeled as geometric clusters, with each cluster containing multiple leaf shapes to add variability. For the stems, a set of curves with variable noise parameters was used to create randomized shapes, simulating natural stem growth. Lighting within the scene was achieved through the ambient illumination provided by the Aigen’s Synthetic Plants Cut Robot Training to About a Week default forest environment texture, which mimicked natural light conditions. In this study, Blender version 4.0.2 was employed to generate the synthetic dataset of strawberry plants for computer vision training. The trained model was subsequently tested on real-world data from the Strawberry Digital Images (StrawDI) dataset , demonstrating the potential of synthetic data and domain adaptation techniques in improving agricultural practices. This study aims to train a ViT-based segmentation model to differentiate between ripe and unripe strawberries using synthetic data generated with Blender.

synthetic data for agricultural robotics

1 Case studies showing yield improvements

This study utilized a Blender-generated synthetic dataset containing 10,000 images, each containing both ripe and unripe strawberries and paired with a corresponding mask. This model is well-suited for handling the complexities of strawberry ripeness detection task due to its ability to capture multi-scale features and fine details. For training, the model leverages a combination of cross-entropy loss and Dice loss to enhance pixel-wise classification and account for class imbalance. The SwinUNet in this study adopts multi-head self-attention and multi-layer perceptron (MLP) layers in its Swin blocks, with ReLU as the activation function. As shown in Fig 5, adapted from Cao et al. , the encoder consists of Swin Transformer blocks, which use shifted windows to efficiently model long-range dependencies in images. After generating the synthetic AI RMF development images using https://aicurated.com/ Blender, the dataset was then divided into training and validation subsets.

  • Your system stops because it can’t classify the objects it’s seeing.
  • A wide realistic multi-crops dataset is accurately generated to train deep semantic segmentation networks and enabling robust performance in challenging real-world conditions.
  • An automatic parametric approach enables an easy customization of virtual field geometry and features for a fast reliable evaluation of navigation algorithms.
  • The datasets are organized by task—ranging from image classification to robotic navigation—and span different sensing modalities, including RGB, multispectral, hyperspectral, and 3D LiDAR.
  • These datasets support bounding box annotations for tasks like fruit counting, weed identification, or tracking crop objects across frames.
  • However, the weather, illumination, uneven terrain, and plants present a number of specific problems for autonomous navigation in agricultural scenarios.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *