LENS: Learning to Navigate with Active Search for Partially Observable MAPF in Unknown Environments

Abstract:

Classical Multi-Agent Path Finding (MAPF) solvers guarantee collision-free coordination but rely on perfect global knowledge, limiting their applicability in strictly unknown environments. Consequently, modern learning-based approaches face a dichotomy: decentralized reactive heuristics scale under partial observability but fail at structured deadlocks due to limited horizons and weak interaction inductive biases, while neural foundation models (e.g., MAPF-GPT) provide topological awareness but require pre-computed global heuristics and prohibitive training data.

We address Centralized Collaborative Partially Observable MAPF (PO-MAPF) by proposing LENS (\textbf{LE}arning to \textbf{N}avigate with active \textbf{S}earch), a hybrid architecture decoupling topological guidance from local collision avoidance. LENS employs a lightweight neural network that maps each agent's field of view (FOV) and goal direction into a dense local potential field to generate multi-step subpaths. Upon anticipating collisions, agents aggregate local observations into a shared belief, projecting neural proposal endpoints as waypoints. LENS then partitions anticipated conflicts into disjoint graphs and invokes localized Conflict-Based Search (L-CBS) over bounded spatio-temporal windows. This ensures strictly collision-free execution within a receding horizon, overcoming reactive myopia without the exponential overhead of global replanning.

Evaluations show that given full global knowledge, LENS approximates the solution quality of centralized oracles across most benchmarks, and achieves out-of-distribution generalization comparable to 85M-parameter foundation models using $<0.2\%$ of their training data. In strictly unknown environments with online mapping, LENS outperforms reactive baselines, improving success rates by $41.5\%$ in high-density mazes. By decoupling topological inference from collision resolution, LENS provides a scalable, data-efficient, and locally collision-free solution for autonomous navigation.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Harsha_Kokel1

Submission Number: 7871