Hex Nut 6D Pose Trainer
Budget: $30 – $250 USD
I need a compact, real-time 6D pose estimation pipeline for a hex nut that my lab robot can grasp reliably, no matter its orientation on the workbench. My current plan is to pair YOLOv12nseg for segmentation with LessAAE for the pose head, then wrap the whole stack in a ROS node that streams camera images, runs inference, and publishes geometry_msgs/Pose messages.
The scene is a controlled laboratory cell, so lighting and background are stable;
Deliverables
• Training scripts and documented config files for YOLOv12nseg + LessAAE
• Trained weights ready for inference on a single hex-nut class
• ROS package with launch file, subscriber/publisher topics, and README for quick deployment
• Short report summarizing dataset setup, hyper-parameters, and achieved mAP/ADD-S metrics
If you have proven experience combining modern YOLO variants with pose-estimation heads and can hand over a clean, catkin-friendly node, I’m ready to start right away.
The scene is a controlled laboratory cell, so lighting and background are stable;
Deliverables
• Training scripts and documented config files for YOLOv12nseg + LessAAE
• Trained weights ready for inference on a single hex-nut class
• ROS package with launch file, subscriber/publisher topics, and README for quick deployment
• Short report summarizing dataset setup, hyper-parameters, and achieved mAP/ADD-S metrics
If you have proven experience combining modern YOLO variants with pose-estimation heads and can hand over a clean, catkin-friendly node, I’m ready to start right away.