ANN-CMCGS: Generalizing Continuous Monte-Carlo Graph Search with Approximate Nearest Neighbors

Christoph Scherer, Wolfgang Hönig

Keywords: Monte-Carlo Graph Search, MCGS, Monte-Carlo Tree Search, MCTS, Robotics, Motion Planning, Online Planning, Approximate Nearest Neighbor, ANN

Abstract: Robot motion planning under uncertainty highlights the need for decision making in continuous domains. Continuous Monte-Carlo Graph Search (CMCGS) meets this need by combining sampling with graph-based reasoning in a unified framework. Existing formulations, however, rely on a layered structure that restricts generality and prevents MCGS from fully realizing its advantages over MCTS. We introduce a non-layered formulation of CMCGS that removes this restriction, enabling a more flexible and scalable search framework applicable to arbitrary directed graphs. To ensure efficiency, we integrate Approximate Nearest Neighbor (ANN) search via Hierarchical Navigable Small-World graphs (HNSW), which allows rapid graph maintenance and querying in high-dimensional continuous spaces. Unlike prior methods that rebuild the graph at every iteration, our approach incrementally updates the graph, bootstrapping the search over time. We demonstrate the benefits in robot motion planning, where non-layered CMCGS with ANN achieves higher success rate compared to layered CMCGS. These results highlight a practical foundation towards real-time, online continuous planning under uncertainty.

Area: Robotics and Control (ROBOT)

Generative A I: I acknowledge that I have read and will follow this policy.

Submission Number: 1401