6IMPOSE: bridging the reality gap in 6D pose estimation for robotic grasping

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• 6D pose recognition is vital for robotic grasping.
• Current deep learning models struggle with real-world generalization.

Purpose of the Study:

• Introduce 6IMPOSE, a novel framework for sim-to-real data generation and 6D pose estimation.
• Enhance the generalization capabilities of 6D pose estimation models for real-world robotic applications.

Main Methods:

• Utilized Blender for synthetic RGBD image dataset generation with 6D pose annotations.
• Developed a real-time, two-stage 6D pose estimation approach integrating YOLO-V4 and a streamlined PVN3D.
• Implemented domain randomization and fine-tuned data generation for improved sim-to-real transfer.

Main Results:

• Generated a synthetic RGBD dataset of five household objects.
• Achieved an 87% success rate in grasping diverse household objects from cluttered scenes.
• Demonstrated successful transfer of trained models to robotic grasping experiments.

Conclusions:

• The 6IMPOSE framework effectively bridges the sim-to-real gap for 6D pose estimation.
• Optimized inference pipeline and data generation overcome generalization limitations of existing algorithms.
• The developed codebase, dataset, and models are publicly available to facilitate further research.