PhD Defense Announcement: Fadhil Ginting
Reasoning and Planning using Contextual Semantics

Zoom: https://bit.ly/GintingDefense

Meeting ID: 705 393 9152 | Passcode: 450496

To enable embodied agents to operate effectively in real-world unstructured environments, the agents must reason about their tasks, surroundings, and embodiment capabilities while planning under partial observability and perceptual uncertainty. Achieving such capability requires grounding reasoning and planning in semantic and contextual knowledge, enabling agents to understand the world more accurately and to plan more efficiently in complex settings.

In this talk, I will present my recent contributions that advance the capabilities of reactive and proactive reasoning and planning through contextual semantics. First, I introduce an approach called SEEK that leverages relational semantic knowledge to perform probabilistic planning in scene graphs, improving object-goal navigation efficiency in large-scale environments. Second, I present a new task, long-term active embodied question answering (LA-EQA), which requires agents to recall past experiences from long-term memory while actively exploring their environments. To address this, I propose a robotic mind palace that enables targeted memory retrieval and guided navigation. Finally, I present SayComply, a framework that grounds robotic task planning in contextual knowledge through retrieval-augmented language models. These contributions are validated in both simulation and real-world robotic deployments, demonstrating their efficacy and practicality for robotic inspection tasks.