Learning Task-Oriented Grasping from Human Activity Datasets
This work addresses the challenge of task-specific grasping for robots, though it appears incremental as it builds on existing methods with a novel joint estimation approach.
The paper tackles the problem of teaching robots task-oriented grasping by leveraging a real-world human activity RGB dataset, resulting in a model that jointly estimates hand and object poses to improve accuracy and enables robots to perform grasping on novel objects with competitive results.
We propose to leverage a real-world, human activity RGB dataset to teach a robot Task-Oriented Grasping (TOG). We develop a model that takes as input an RGB image and outputs a hand pose and configuration as well as an object pose and a shape. We follow the insight that jointly estimating hand and object poses increases accuracy compared to estimating these quantities independently of each other. Given the trained model, we process an RGB dataset to automatically obtain the data to train a TOG model. This model takes as input an object point cloud and outputs a suitable region for task-specific grasping. Our ablation study shows that training an object pose predictor with the hand pose information (and vice versa) is better than training without this information. Furthermore, our results on a real-world dataset show the applicability and competitiveness of our method over state-of-the-art. Experiments with a robot demonstrate that our method can allow a robot to preform TOG on novel objects.