Doctoral Thesis: Neural Fields for Robotic Manipulation

Yen-Chen Lin

Abstract:
Neural fields, or coordinate-based neural networks, have garnered significant attention for their ability to parametrize physical properties across space and time, thereby offering novel solutions to complex visual computing challenges. This talk studies applications of neural fields within the domain of robotics. First, we demonstrate how neural fields can be employed to synthesize data, generating rich and varied training samples that enhance the robustness and accuracy of machine learning models for robotic perception. Second, we present a novel application of using neural fields to simulate camera properties and show that its ability to generate orthographic views of objects and environments is critical for affordance-based robotic manipulation systems. Through the precise prediction of orthographic views, robots are empowered to perform 6-DoF manipulation tasks without relying on prone-to-error depth sensors. Lastly, we investigate the scalability of neural fields for city-scale modeling, providing a comprehensive framework that integrates modularity and acceleration data structures such as hash grids and bit fields. This large-scale modeling approach has the potential to drive advancements in autonomous navigation, urban planning, and informed decision-making for robotic deployments in urban environments. Collectively, these projects underscore the versatility and value of neural fields as cutting-edge computational tools for advancing the field of robotics.

Committee:
Phillip Isola, Alberto Rodriguez, Vincent Stizmann

Details

• Date: Friday, June 16
• Time: 3:00 pm - 4:30 pm
• Category: Thesis Defense
• Location: 34-401 (Grier Room)

Additional Location Details:

To attend via Zoom:

https://mit.zoom.us/j/99846803422?pwd=RGgvbVdxMVpsOWRMSVRYSUhsNHVkUT09

Password: 104401