Technical Summary: UGV-Conditioned Multi-UAV Informative Planning on a Shared Exposure Belief

Problem Statement

In large-scale, threat-augmented environments, safe ground navigation for Unmanned Ground Vehicles (UGVs) requires active aerial support to mitigate risks. Existing aerial reconnaissance systems typically focus on generic mapping, terrain coverage, or traversability estimation. However, these approaches often fail to direct sensing toward regions most critical for ground vehicle safety. Specifically, current systems lack a closed-loop mechanism where cooperative multi-UAV informative planning is tightly coupled with exposure-aware UGV routing. There is no shared environment belief that simultaneously drives UAV sensing priorities and adapts UGV path selection as new information arrives online, leading to redundant sensing and suboptimal safety outcomes.

Methodology

The authors propose a framework where a team of $N$ UAVs coordinates to support a single UGV navigating through unknown threat zones. The core of the approach is a shared exposure belief that is updated online from aerial observations and utilized by both the UGV planner and the UAV team.

1. Shared Exposure Belief

The system maintains a probabilistic map of the environment's exposure field, discretized into a grid.

• Modeling: The latent exposure rate $\lambda(x)$ at each cell is modeled as a log-normal random variable. Initial estimates are derived from terrain-class priors (e.g., roads, forests), while epistemic uncertainty is captured by the variance.
• Update Mechanism: As UAVs observe cells, the belief is updated using a Bernoulli-Shannon entropy metric. Cumulative sensor coverage reduces the effective log-standard deviation (uncertainty) and updates the mean. Detected threat zones locally amplify the belief via a multiplicative factor.
• Risk-Averse Surrogate: Both planners query a scalar risk-averse surrogate $\lambda_{risk}$, which combines the expected exposure value with a penalty for high uncertainty. This ensures that unobserved cells are treated as potentially hazardous, incentivizing reconnaissance.

2. Exposure-Aware UGV Replanning

The UGV planner seeks a path $\pi$ from start to goal that minimizes a cost function comprising traversal time and cumulative exposure:
$$\pi^* = \arg\min_{\pi} [T(\pi) + \beta E(\pi, t)]$$
where $E(\pi, t)$ is the cumulative exposure calculated using the risk-averse surrogate. The UGV replans asynchronously whenever a UAV detects a new threat zone, committing to a fixed-length trajectory prefix to ensure safe execution while the global path is recomputed.

3. UGV-Conditioned Informative Planning (UAV Objective)

Unlike standard Informative Path Planning (IPP) that maximizes generic information gain, the UAVs optimize a mission-conditioned reward. The reward function $R(x, t)$ for a cell $x$ is constructed to prioritize observations that impact the current UGV route:

• Corridor Proximity: A gating mechanism suppresses rewards for cells far from the current UGV route, ensuring sensing effort is focused on mission-relevant regions.
• Exposure Modulation: Within the corridor, the reward is scaled by an exposure-dependent gain. High-exposure or high-uncertainty regions near the route receive higher priority.
• Entropy: Uncertainty (entropy) modulates the urgency of observation within the gated region.
  This approach directs UAVs to observe "route-relevant, uncertain, and high-exposure" regions rather than performing generic exploration.

4. Multi-UAV Region-Conditioned Coordination

To prevent multiple UAVs from converging on the same high-reward corridor regions (redundant sensing), the framework employs spatial ownership:

• Voronoi Assignment: The environment is partitioned into disjoint regions assigned to each UAV based on travel cost to the region centroid.
• Reward Masking: Each UAV's effective reward is masked by its assigned region.
  • Hard Coordination: UAVs perceive zero reward outside their assigned region.
  • Soft Coordination: UAVs retain a fraction of the reward from foreign regions, allowing responsiveness to high-value threats emerging near region boundaries.
    This coordination layer reduces redundancy while maintaining scalability and responsiveness to dynamic changes in the UGV's route.

Key Contributions

The paper makes three primary claims and contributions:

1. Shared Exposure Belief: Coupling UAV sensing actions to a shared, exposure-aware belief significantly improves downstream UGV safety compared to planning based solely on prior terrain information.
2. Route-Conditioned Sensing: Conditioning the UAV objective on the current UGV route focuses sensing effort on mission-relevant regions, rather than generic exploration or uniform uncertainty reduction.
3. Region-Conditioned Coordination: In multi-UAV settings, assigning spatial ownership reduces redundant sensing while preserving responsiveness to newly relevant areas, addressing the scalability challenges of cooperative informative planning.

Experimental Results

The approach was validated through 100 Monte Carlo simulation trials in a 5 km $\times$ 5 km environment derived from OpenStreetMap data, featuring unknown threat zones.

• Safety Improvement: Compared to a system using terrain-semantic replanning without exposure modeling, the proposed exposure-aware approach reduced cumulative UGV exposure by 38%. It also reduced the mission objective cost by 31% relative to the terrain-semantic baseline.
• Reward Architecture: Ablation studies showed that the proposed gated reward architecture outperformed additive, multiplicative, and weighted entropy baselines. It achieved the lowest mission cost while maintaining high corridor coverage and priority-threat detection rates.
• Coordination Efficiency: Multi-UAV coordination significantly reduced redundant coverage.
  • Without coordination, redundant coverage was 38.8%.
  • With soft-regional coordination, this dropped to 5.0%.
  • With hard-regional coordination, it dropped to 3.7%.
  • Useful information efficiency increased from 66.2% (no coordination) to over 92% (soft/hard regional).
• Scalability: As the number of UAVs increased (up to 6), region-conditioned masking prevented the rapid growth of redundant sensing seen in uncoordinated teams, though strict hard masking showed diminishing returns in threat detection time for larger teams due to region fragmentation.

Significance and Claims

The paper claims that its primary significance lies in closing the loop between cooperative multi-UAV informative planning and exposure-aware UGV routing. By maintaining a shared belief that drives both the ground vehicle's path selection and the aerial team's sensing priorities, the system enables safer navigation in large-scale, partially observed environments.

The authors emphasize that their approach moves beyond generic exploration or mapping; the value of aerial sensing is explicitly defined by its expected impact on downstream UGV safety. While the current evaluation assumes idealized sensing and instantaneous communication, the results demonstrate that the planning and coordination mechanisms themselves are effective at reducing exposure and eliminating redundancy. The work establishes a foundation for future hardware validation and more realistic sensing models.