Preview

Description

Robots will become more prevalent in industry and our everyday lives as we continue on the current trend of automation. In a variety of settings, robots need to robustly interact with their environment to successfully accomplish various tasks. Towards this goal of robust robotic systems, I have worked on two projects during the course of my master's studies.

The first project is Mechanical Search, a class of tasks that requires a robot to locate and extract a target object. We implement a physical system for a particular instance of the Mechanical Search problem that involves retrieving a target object from a heap of objects in a bin by leveraging recent advancements in computer vision and using action primitives such as grasping and pushing. For this project, I worked on improving and evaluating the target recognition segment of the pipeline through experiments with varying Siamese network architectures and dataset augmentation techniques.

The second project is Adversarial Grasp Objects, in which we explore an analog of adversarial images in the domain of robust robot grasping to synthesize objects that are difficult for a robot to grasp, but appear similar in shape to existing objects. By doing so, we can analyze the failure modes of a grasp planner. We explore two algorithms for generating such objects: an analytical algorithm and a deep learning algorithm. For this project, I developed one variant of the analytical algorithm that minimally perturbs vertices on antipodal faces in randomly sampled directions subject to geometric constraints to maintain similarity to the input object. I also conducted experiments and analyses on the objects generated by both types of algorithms and evaluated adversarial grasp objects on a physical system.

Details

Files

Actions

Statistics

from
to
Export
Download Full History