The problem of object shape and pose estimation is integral to various robotic tasks ranging from manipulation to object-level localization and mapping. Simply put, it is often important for robots to know where objects are and how they look. Given that only partial observations are often available due to self-occlusions or occlusions in the scene, past works have attempted to learn object shapes from incomplete data via vast training datasets. While effective for previously known and well defined objects (e.g., cars and bottles), these methods could fall short when estimating shapes of arbitrary objects (e.g. rocks or deformed packages). In this work we present a novel approach to estimating continuous and differentiable shape-estimates to partially observable objects. Leveraging past work on computing ellipsoid shape estimates to objects, our method improves on such coarse estimates by deforming the prior ellipsoids to tightly fit partial observations while retaining a reasonable volume and without relying on prior shape or semantic knowledge.

Gaining experience signal processing and estimation (6.011), algorithms design (6.046) and hands on work with learned perception systems (at RRG) got me very excited about putting those skills to the test through SuperUROP research! Having worked with deep neural networks and classical deterministic algorithms before, I look forward to be working with uncertainty models which are very different in nature; both white-boxed and stochastic.