Markov decision process (MDP) planners typically rely on a nominal transition model that may not match the true dynamics. Robust MDP planning handles this uncertainty by considering a set of admissible models, but doing so under tight online computation limits has been a challenge.
The new work introduces a scalable robust Monte Carlo tree search (MCTS) method that maintains anytime deterministic optimality guarantees. This means the algorithm can return a solution at any point during computation, with a guarantee that improves as more time is allowed, while still scaling to large problems.
The abstract is brief, but the core contribution appears to be a practical bridge between robust planning and MCTS efficiency. The authors emphasize both scalability and formal guarantees, which are often in tension in online planning. Further details are not available from the source, so the exact algorithmic mechanisms remain unspecified.