Scalable Rao-Blackwellized Online Planning for High-Dimensional POMDPs
A paper extends the Rao-Blackwellized online POMDP framework to high-dimensional settings using hybrid continuous-discrete belief representations. It integrates with FastSLAM 2.0 in a robotic search-and-rescue task and achieves higher cumulative rewards with fewer particles and simulations than purely sampling-based methods.
The paper presents an extension of the Rao-Blackwellized online POMDP (RB-POMDP) framework for high-dimensional partially observable environments. It uses hybrid continuous-discrete belief representations to analytically propagate uncertainty of marginalized state components during tree-based planning, reducing variance in value estimation. In a robotic search-and-rescue task integrated with FastSLAM 2.0, the planner achieves higher cumulative rewards using significantly fewer particles and planning simulations than purely sampling-based methods under equivalent computational budgets.
The approach reduces Monte Carlo variance by analytically marginalizing continuous state components, enabling more efficient belief propagation in tree search. Integration with FastSLAM 2.0 suggests compatibility with SLAM-based state estimation, potentially improving planning accuracy in high-dimensional robotic tasks.
Robotic systems in search-and-rescue and other partially observable environments may benefit from more sample-efficient planning, reducing computational requirements and enabling real-time decision-making in complex settings.
Improved planning efficiency could lower hardware and compute costs for autonomous robots, making advanced POMDP-based decision-making more accessible for commercial applications such as logistics, inspection, and emergency response.
Further validation in diverse robotic domains and comparison with other state-of-the-art POMDP solvers would clarify generalizability. Potential extensions include handling more complex observation models and scaling to multi-robot coordination.