Defense of doctoral thesis “Towards Robust 6-DoF Multi-Finger Grasping in Clutter with Explicit Scene Understanding”, M.Sc.(Tech.) Jens Lundell
M.Sc.(Tech.) Jens Lundell will defend the thesis “Towards Robust 6-DoF Multi-Finger Grasping in Clutter with Explicit Scene Understanding” on 25 February 2022 at 12 in Aalto University School of Electrical Engineering, Department of Electrical Engineering and Automation.
When? 25.02.2022 12:00 – 15:00
Where? Remotely. Zoom link: https://aalto.zoom.us/j/64400197025
Opponent: Prof. Danica Kragic, KTH, Sweden
Supervisor: Prof. Ville Kyrki, Aalto University School of Electrical Engineering, Department of Electrical Engineering and Automation
Thesis available for public display at: https://aaltodoc.aalto.fi/handle/123456789/112725
For robots to become widely adopted in everyday human lives, they need the ability to quickly pick a wide variety of everyday objects such as cutlery, toys, and food without dropping them. Although much progress has been made towards this goal in the last decade, four fundamental robotic grasping problems are still unsolved. These are robust grasping, 6-degree-of-freedom grasping, multi-finger grasping, and grasping in clutter. In this dissertation, I propose building grasp samplers that use explicit scene understanding to solve these problems.
In the dissertation, I define explicit scene understanding as reconstructing the shape of all partially observed objects in the scene. To do the actual scene reconstruction, I used a machine learning technique called deep learning. With this technique, I developed a general reconstruction method that could complete the shape of partially observable objects and reason about the reconstruction uncertainty. Furthermore, I presented four different grasp planners that use this reconstruction method to sample grasps: one for sampling grasps robust to shape uncertainty, one for sampling six-degree-of-freedom grasps, one for sampling multi-finger grasps on singulated objects, and, finally, one for sampling multi-finger grasps on objects in clutter.
Together, all the results presented in my dissertation indicate that explicit scene understanding increases the generality, robustness, and performance of 6-degree-of-freedom parallel jaw and multi-finger grasp samplers, albeit at a higher computational cost. Especially the multi-finger grasp samplers are among the first to achieve sub-second grasp sampling, which is a significant improvement over previous state-of-the-art multi-finger grasp samplers. Consequently, if the extra computational demand for object reconstruction is not of concern, we recommend that roboticists consider explicit scene understanding when developing new grasping approaches.
More details: Aalto events