Undergraduate

Previous Projects

MAPF problems are typically defined as a one-shot problem where a query is provided to the algorithm and it is tasked with computing a collision-free path for a fleet of agents between their respective start and target locations. However, in many cases the environment is known in advance (e.g., in warehouses) and the system is tasked with answering multiple queries. While this layout is constantly changing, at any given point of time, the vast majority of the environment remains unchanged when compared to the original layout.

From a minimally-invasive medical robot automatically inspecting the surface of a diseased organ to an autonomous quadcopter inspecting the structure of a bridge, robots in a variety of scenarios must plan their motions to efficiently inspect regions of interest. These are examples of robot-inspection planning tasks which arise in diverse applications. Here, we are tasked with planning motions for a robot that enable it to efficiently inspect a region of interest using its on-board sensors. Such an inspection plan should not only inspect the region of interest but also obey the robot's kinematic constraints, avoid obstacles (e.g., parts of a bridge or sensitive anatomical structures in bridge inspection and medical applications, respectively), and minimize some metric (such as time to completion or distance traveled).

As humans, the placement and directivity of our ears give us the notion of the spatial layout of a scene that we observe. Sound emitting objects on our right will elicit louder response in the right ear, and similarly with objects on our left. Generating such an effect using loudspeakers or headphones, substantially enhances the user experience by creating a 3D soundsensation. Traditional sound spatialisation requires expensive equipment and nontrivial expertise. This project implements a system which transforms a normal field of view video with a mono recording to a binaural recording using a neural network.

Filtering out interfering sound signals, a procedure known as source separation has been a longstanding problem in the speech processing community. The visual modality is strongly correlated with its accompanied audio. For example, lip reading and motion trajectories can be exploited to improve the separation quality. In this project, a source separation technique for e.g. musical instruments was implemented. A weakly supervised approach for source separation was utilised. A pretrained ResNet visual branch first provides weak labels for the objects in the video. On the audio side, the non-negative matrix factorization (NMF) decomposition of the mixture signal is computed, resulting in a learnt dictionary matrix for the magnitude spectrum. The sources are separated by assigning the basis vectorsto the objects defined by the visual labels.

From a minimally-invasive medical robot automatically inspecting the surface of a diseased organ to an autonomous quadcopter inspecting the structure of a bridge, robots in a variety of scenarios must plan their motions to efficiently inspect regions of interest. These are examples of robot-inspection planning tasks which arise in diverse applications. Here, we are tasked with planning motions for a robot that enable it to efficiently inspect a region of interest using its on-board sensors. Such an inspection plan should not only inspect the region of interest but also obey the robot's kinematic constraints, avoid obstacles (e.g., parts of a bridge or sensitive anatomical structures in bridge inspection and medical applications, respectively), and minimize some metric (such as time to completion or distance traveled).

MAPF problems are typically defined as a one-shot problem where a query is provided to the algorithm and it is tasked with computing a collision-free path for a fleet of agents between their respective start and target locations. However, in many cases the environment is known in advance (e.g., in warehouses) and the system is tasked with answering multiple queries. While this layout is constantly changing, at any given point of time, the vast majority of the environment remains unchanged when compared to the original layout.