Manipulation robotics
CNN Grasp Detection
Learning grasp position, angle, and width from depth images for an ABB YuMi manipulator.
M.Sc. thesis · Bauman Moscow State Technical University
The problem
Robotic pick-and-place requires converting visual observations into grasp parameters that a manipulator can execute.
My contribution
- Trained a CNN on depth images to predict grasp position, angle, and width.
- Connected grasp predictions to a pick-and-place pipeline.
- Validated the approach in CoppeliaSim with the ABB YuMi dual-arm manipulator.
Technical approach
Depth observations feed a learned grasp detector. Its predicted grasp parameters inform the manipulation pipeline for simulated execution.
Results & validation
Perception-to-action pipeline on ABB YuMi
The system was evaluated on test objects. The evaluation material will specify the success metric, test set, and experimental conditions.
Depth input and grasp predictions
