Generalizable Long-Horizon Robotic Manipulation under Uncertainty and Partial Observability

Abstract

A central goal in embodied artificial intelligence is to enable autonomous agents to accomplish complex, long-horizon tasks in novel, partially observable environments. In these scenarios, agents must effectively reason about uncertainty, generalize from limited experiences, and proactively plan actions to acquire missing information. This thesis tackles these core challenges by developing and evaluating novel methods specifically designed for partially observable contexts. The first part of this thesis introduces an enhanced heuristicguided planning technique that increases search efficiency in sparse-reward domains with significant uncertainty. Next, we investigate how symbolic reasoning can be integrated into the decision-making framework, accelerating search through the use of temporal and belief-space abstractions. Next, we propose a method for sequencing low-level reinforcement learning skills alongside information gathering actions, enabling increased task complexity and robustness in real-world tasks. Lastly, we show how large language models may be leveraged for few-shot model learning, allowing agents to rapidly adapt and generalize to new scenarios. The methods presented in this thesis advance the state-of-the-art in embodied AI by enabling robots to better handle uncertainty and incomplete information, ultimately paving the way for more capable, exploratory, and risk-aware autonomous systems.