HTP Graphics

RAIN PROGRESS // We have developed two models based on variational autoencoders which generates multiple grasp maps per input image. We first compared them against previous models by evaluation on the Cornell and Jacquard datasets, which are the commonly used datasets for 2D grasp evaluation and have achieved good results so far. For real-world tests, the initial implementation was completed around the end of last year but due to the lockdown, we are now using the simulator platform to first refine and evaluate our models before applying it in real world. For this, we integrated the grasping model with the virtual platform and are applying it on a variety of objects while controlling the level of clutter in the environment. We are currently evaluating with 3D models from the Evolved Grasping Analysis Dataset (EGAD) and Dexnet Adversarial dataset which contain objects with a large range of shape complexity and grasp difficulty. After this evaluation stage has been completed, the next step will be to reproduce the results with the actual robot. REMOTE HANDLING 51 FUTURE ASPIRATIONS // Data driven techniques and especially deep learning has a lot of potential in addressing many of the challenges of robotic grasping such dynamic environments, adapting to unknown objects and gripper types. Due to the limited availability of labelled grasp data (which could be in the form of oriented rectangles representing the position, angle and width gripper for a 2D image or a gripper configuration associated with a 3D mesh or pointcloud), especially for clutter and non-standard object shapes, one of the possible directions of our research could be to use the simulated grasp runs for learning and generating datasets. There is a lot of interesting research that can be explored further within the field of robotic grasping, as even with the incredible number of advancements in recent years, we are still very far from achieving the ideal dexterity of the human hand.

• Cover