Object-Centric Planning for Long-Horizon Robotic Manipulation and Navigation

Abstract

A primary objective within the robotics research community is the development of robotic agents capable of executing long-horizon tasks within complex and novel environments. The sparse and factored nature of object-centric planning makes it a good candidate for the reasoning engine inside such an agent. However, several challenges remain under an object-centric planning framework. Challenges arise in areas such as efficiently grounding states with novel objects in cluttered environments, maintaining efficiency under large object sets, and safe exploration and manipulation in partially observable and nondeterministic environments. This thesis examines these limitations and proposes several strategies for solving them while maintaining the generalizability and flexibility of object-centric planning in long-horizon tasks.