Non-Prehensile Mobile Agent Interactive Navigation in Cluttered Environments

Abstract

As autonomous mobile robots are increasingly deployed in complex, real-world environments, traditional approaches that rely solely on collision-free navigation often fall short. In many practical scenarios, physical interactions with the environment, such as pushing objects or manipulating articulated structures, are not only unavoidable but also crucial for completing tasks. This thesis studies non-prehensile interactive navigation (NPIN), where robots interact with their surroundings using non-grasping, contact-based actions. These interactions present significant challenges, including complex interaction dynamics, long-horizon planning under uncertainty, and the lack of standardized tools for evaluation.

This thesis addresses these challenges through three core contributions. First, we present a novel learning-based predictive planning algorithm for autonomous surface vehicle (ASV) navigation in ice-covered waters, a real-world instantiation of NPIN. In this scenario, the ASV must interact with dynamic ice floes to reach its goal. We propose a hybrid planning framework that integrates deep learning-based occupancy prediction of obstacle motion with a graph search-based planner. The resulting system is capable of computing safe and efficient trajectories that account for future ice movements, both in simulation and in a physical testbed.

Second, we broaden the scope and introduce Bench-NPIN, the first standardized benchmarking suite for NPIN. Bench-NPIN includes a diverse set of environments and tasks, such as maze navigation, box delivery, and area clearing, which span both navigation-centric and manipulation-centric categories. It also provides unified evaluation metrics and reference baseline algorithms to facilitate fair and reproducible comparisons across different approaches.

Third, we present a forward-looking study into long-horizon planning through generative skill chaining using diffusion models. We investigate how sequences of low-level interaction skills can be generated in one shot using a learned generative model. Furthermore, we introduce an out-of-distribution (OOD) detection mechanism to evaluate the feasibility of these skill sequences. This allows the planner to anticipate failure, reject infeasible plans, and potentially replan in a timely manner.

Together, these studies provide a comprehensive investigation into the foundations, applications, and emerging approaches for NPIN in cluttered and uncertain environments.