Task and motion planning for mobile manipulation

Thesis

Task and Motion Planning (TAMP) aims to integrate high-level symbolic reasoning with low-level geometric reasoning, creating extensive plans grounded in actionable commands. By bridging decision-making and physical action, TAMP enables robots to perform complex, multi-step tasks.

This thesis contributes solutions to the TAMP problem for mobile manipulation through three distinct projects, which are designed to exploit unique challenges of TAMP. The frameworks address embodied intelligence through investigation across sampling, optimization, and learning methods.

First, a hierarchical optimization-based TAMP structure is introduced for coverage planning in legged manipulators, enabling efficient and robust execution with a dexterous task. Second, an approach combining sampling-based methods with optimal TAMP principles is developed, constructing a reachability graph for efficient state-space exploration. Finally, an LLM-based planner is designed to enable natural language interaction, lightweight adaptability, and failure recovery in domestic environments. These systems are validated via different robot deployments in industrial and domestic environments, aiming at robust and reliable robot autonomy.

In addition to solving TAMP, our frameworks design end-to-end solutions by processing perception input to the central planner, further contributing towards the autonomy and robustness of robotics systems.

Authors: Ly, KT

• DOI: 10.5287/ora-r8xmwgjjd
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: mobile manipulation; artificial intelligence; task and motion planning; embodied AI
• Subjects: Robotics