Bin-Picking Model for 6DoF Pose Estimation

Info "At a glance"

Goal: Estimate 6-DoF pose of known parts for bin-picking, so a robot can place a grasp in SE(3) despite clutter and occlusion.
Approach: Detect/segment the target, align the CAD model to observations (keypoints/contours → PnP), then refine with depth/ICP and reject outliers by score/visibility.
Inputs/Outputs: RGB-D frames → object pose T_world^obj (+ grasp pose) with confidence.
What worked: CAD-guided refinement and depth-based filtering improved stability under partial occlusion; symmetric-object handling reduced pose flips.
Failure modes: Heavy occlusion, specular/texture-poor parts, and extreme bin clutter; runtime depends on detector & ICP settings.
Next: Multi-view fusion, uncertainty-aware grasp sampling, and on-robot validation.

If you'd like to do a deep dive further on the project, Please feel free to go through the complete write-up below.

Full report(PDF) is embedded below. Download links are provided if your browser blocks inline PDFs.