📝 Abstract:
To create a map of unknown terrain where GPS is unreliable or unavailable, decentral ized UAV swarms are deployed. This task demands a trade-off between maximizing spatial coverage, maintaining network connectivity, and limiting the localization drift accumulated by the LiDAR sensors on UAVs. Existing work treats connectivity and drift correction as fixed constraints instead of coupled behaviors. The purpose of this thesis is to develop a decentralized Multi-Agent Reinforcement Learning (MARL) model to solve this problem. Each agent learns to cover unexplored regions, maintain network connectivity by dynamic relay allocation, and show drift-aware behaviors together. Agents manage drift by implicitly performing loop closures as localization confidence decreases. The MARL policy is trained with a curriculum learning paradigm, where the scale of the mission area and UAV count increase over stages. It is evaluated using a four way ablation design that isolates the individual contribution of FANET-awareness and SLAM-drift-awareness. This is achieved by toggling each constraint in direct link, single-hop, and multi-hop scenarios. The policy is evaluated on the same three v scenarios with different swarm sizes. Results show that under direct-link conditions, the policy matches the spatial coverage of a drift-and-connectivity-unaware baseline to within 0.1 percentage points (95.6% vs. 95.7%). As the scale of the mission area grows, relay usage and multi-hop path length also grow proportionally. For example, relay ratio rises from 4.6% to 33.0%, and the 2+-Hop Ratio rises from 11.5% to 55.7%. A notable observation is that managing connectivity and drift together does not affect the overall spatial coverage considerably. In the cases requiring multi-hop paths, the policy and the drift-unaware baseline demonstrate approximately the same coverage and relay workloads. Results also show that the number of loop-closures and the mission area scale are inversely proportional. The loop-closure rate drops to zero in multi-hop scenarios, but localization confidence remains bounded. A swarm-size scaling analysis was also performed. This revealed that packet delivery reliability increases proportionally with the number of agents. Spatial coverage rate drops when swarm size does not scale with the required multi-hop paths. In summary, a policy that balances exploration, network connectivity, and localization confidence without explicit coordination can be learned.