Primary supervisor

Hamid Rezatofighi

Co-supervisors

This project develops PACE-Drone, an intelligent drone planning system that learns from experience rather than following pre-programmed scripts. Unlike current drones that treat each mission independently, PACE-Drone maintains a persistent belief over user preferences via Bayesian learning, actively discovers implicit constraints from historical mission logs, and balances exploration with task completion based on instruction granularity. The system combines POMDP planning, Gaussian Process preference learning, and LLM-based explanation generation to create truly adaptive autonomous drones for warehouse inspection, search and rescue, and infrastructure monitoring.

Aim/outline

  1. Literature review on preference learning, constraint acquisition, and belief space planning.

  2. Implement Bayesian preference learning module using Gaussian Processes to update user preference beliefs across missions.

  3. Develop log mining algorithms to discover spatial, temporal, and contextual patterns from mission history.

  4. Design width-modulated POMDP planner that adapts exploration-exploitation balance based on instruction granularity (fine vs. coarse instructions).

  5. Build LLM-based explanation generator to present top-k plans with preference-aware justifications.

  6. Evaluate in drone simulation (AirSim/(HAMERITT https://arxiv.org/abs/2409.06608)) on warehouse inspection and search scenarios.

URLs/references

NEUSIS: A Compositional Neuro-Symbolic Framework for Autonomous Perception, Reasoning, and Planning in Complex UAV Search Missions https://arxiv.org/abs/2409.10196

IMPORTANT NOTE

For any inquiries regarding this project, please contact Dr. Sukai Huang via sukai.huang@monash.edu

Required knowledge

  • Python programming and PyTorch

  • Reinforcement learning fundamentals (POMDPs, belief state estimation)

  • Probabilistic modeling (Gaussian Processes, Bayesian inference)

  • Basic robotics concepts (path planning, autonomous systems)

  • Large Language Model integration (prompting, structured outputs)