Instead of aiming to develop a silver bullet grasp synthesis algorithm that can generalize over any possible object and condition, our research seeks to integrate the outputs of multiple grasping algorithms to achieve a higher degree of robustness and accuracy compared to the individual algorithms themselves. We achieve this goal via an ensemble learning framework in which the individual grasp synthesis algorithms are treated as “experts” providing their “opinion” for a given input, and a supervisory system trained using convolutional neural networks intelligently combines these opinions and outputs the final decision. We implement this framework by developing a Mixture of Experts (MOE) ensemble model by adopting GQCNN-4.0, GGCNN, and GGCNN-RGB (our custom implementation of GGCNN using RGB images) as experts. We run our MOE model on the Cornell Grasp Dataset, and compared its performance to the individual success rates of the experts. We obtained a 6\% improvement over the best performing individual network, which is achieved via the ability of the MOE model to combine the advantages of each expert, while overcoming their individual disadvantages through learning how to best partition the input space. We also provide experiments using a Franka Emika robot and obtained significant improvements via our ensemble strategy compared to individual expert opinions.
Our ensemble learning scheme
Our related papers in this project:
ECNNs: Ensemble Learning Methods for Improving Planar Grasp Quality Estimation,
F. Alladkani, J. Akl, B. Calli
IEEE International Conference on Robotics and Automation (ICRA), 2021.
[Paper] [Video]