Terrain-Aware Thompson Sampling for Multi-Drone Search and Rescue

arXiv:2611.14685 · cs.RO · Submitted 2026-11-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Terrain-Aware Thompson Sampling for Multi-Drone Search and Rescue".

Dev: Here's the gist of the paper. Terrain-Aware Thompson Sampling for Multi-Drone Search and Rescue proposes a decentralized terrain-aware multidrone Thompson Sampling algorithm tailored for search and rescue (SAR) applications.

Rosa: The proposed method enables a balance between exploration of uncertain regions and exploitation of high-probability areas with the goal of locating a lost person.

Dev: Algorithm Foundation The core decision-making framework utilizes Thompson Sampling, which is a Bayesian decision-making framework increasingly used in search and rescue.

Rosa: Agents sample from a posterior to select actions with the highest sampled reward, naturally balancing exploration and exploitation.

Paper summary: Rosa: The main idea is that you can balance exploring unknown areas with exploiting the spots where you have the best chance of finding someone. They use Thompson Sampling, which is a Bayesian way to make choices where you sample from a probability distribution to pick what looks best right now. Dev So they’re using this sampling process to guide the drones toward promising locations, right?

Taro: But it’s not just random sampling; they're feeding that probability distribution with terrain data, like elevation and how water flows or the path of a lost person. That data comes from historical SAR case information. Rosa Exactly. They define a mean probability for every single spot on the map based on all those features—displacement ring, watershed, track offset—which helps them figure out where a lost person is most likely to be.

Dev: And they aren't just relying on that prior belief; they incorporate observation uncertainty too. They measure how much of an area is blocked by trees using something called fractional forest coverage data from Google Earth Engine, which tells you the percent of an observation that’s hidden by vegetation. Taro So if a spot looks good based on terrain but is completely hidden by a dense forest, the uncertainty increases and that affects the decision-making.

Rosa: Right. And to keep things stable while they're running these updates, they introduce a scaling factor, which we call lambda, to regulate how strong the update is when new observations come in. Dev That sounds like a necessary tweak for real-time systems to make sure it doesn't just chase every little flicker of data.

Taro: They also have this part where agents try not to go everywhere at once. They add a repulsion term into the reward function, which means if one drone picks a spot, the cost goes up for other drones trying to pick that exact same spot along their path. Rosa So it discourages redundant exploration, pushing them to slightly detour toward better spots instead of just repeating what everyone else is doing.

Dev: It’s smart because they are combining these ideas—the Bayesian updating with terrain priors and the uncertainty from forest cover—to aim for faster target localization instead of just spreading out coverage evenly across the whole search area.

Conclusion: Rosa: So, looking at this paper, "Terrain-Aware Thompson Sampling for Multi-Drone Search and Rescue," it seems they’ve put together a system that uses the environment to make smarter drone decisions when searching for someone. Dev It really boils down to taking historical SAR data, mapping it onto terrain features like elevation and watersheds, and then letting drones use Thompson Sampling to decide where to go next based on where they think the person is most likely hiding.

Taro: The implication for us is that instead of just sending drones out in a grid pattern trying to cover every inch, this method prioritizes high-probability regions based on that terrain knowledge, which should lead to finding the person faster than uniform exploration. Rosa It’s about shifting from just being thorough to being targeted efficiently when time matters in a rescue situation.

Dev: The authors tackle the problem of decentralized coordination that usually fails when communication is spotty in SAR zones by using Thompson Sampling as a natural framework for agents to cooperate without needing a central command constantly feeding them instructions. Taro That distributed decision-making aspect, where each drone makes its choice independently but still contributes to the shared goal, is really important for real-world deployment.

Rosa: And they’ve clearly shown how incorporating uncertainty from things like forest coverage helps them account for the fact that not everything they see is actually what it seems in a search environment. Dev Overall, this work takes established Bayesian methods and grounds them in specific physical realities of SAR—terrain and observation limits—to create a more effective way for multi-drone teams to locate missing people.

Polytechnique Montreal

cs.RO

Submitted: 2026-11-10

Updated: 2026-11-10

Comments: Accepted to the 18th International Symposium on Distributed Autonomous Robotic Systems (DARS 2026)

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 89/100

The gist: Terrain-Aware Thompson Sampling for Multi-Drone Search and Rescue proposes a decentralized terrain-aware multidrone Thompson Sampling algorithm tailored for search and rescue (SAR) applications.

Key concepts

Thompson Sampling
A Bayesian decision-making method where agents sample potential outcomes based on current beliefs. This naturally balances exploring new areas (exploration) with focusing on known good areas (exploitation) by selecting actions with the highest sampled reward.
Terrain Prior Modeling
The search map is informed by historical SAR data describing lost person behavior, including elevation and track offsets. These features create a probability distribution for each grid cell, guiding where drones are most likely to find a person.
Fractional Forest Coverage
This measures how much of an observation is hidden by tree cover. This data quantifies how difficult it is to detect ground features in dense vegetation, and this uncertainty is used to adjust the search strategy.

Terminology

Summary

Terrain-Aware Thompson Sampling for Multi-Drone Search and Rescue proposes a decentralized terrain-aware multidrone Thompson Sampling algorithm tailored for search and rescue (SAR) applications. p. 0 The proposed method enables a balance between exploration of uncertain regions and exploitation of high-probability areas with the goal of locating a lost person. p. 0

Algorithm Foundation

The core decision-making framework utilizes Thompson Sampling, which is a Bayesian decision-making framework increasingly used in search and rescue. Agents sample from a posterior to select actions with the highest sampled reward, naturally balancing exploration and exploitation. p. 1 This approach maintains a probabilistic belief over unknown quantities via a prior probability distribution that is updated as observations are collected (posterior). p. 1

Terrain Prior Modeling

The map priors are informed by established SAR models derived from historical SAR case data of lost person behavior. These models include displacement ring, elevation, watershed, dispersion angle, and track offset. p. 0 The mean probability value µ for each grid cell ci is defined as a normalized weighted mixture of the K probability density functions fk(·) derived from the heterogeneous search features introduced above. p. 4

Observation Uncertainty Handling

Observation uncertainty is incorporated through fractional forest coverage, measuring the percent of an observation occluded by tree coverage and accounting for reduced detection performance in densely vegetated areas. The uncertainty σ2ci encodes the level of ground occlusion (%), obtained from fractional forest coverage data from Google Earth Engine. p. 6 A scaling factor λ is introduced to regulate the effective update strength and ensure stable convergence of the inference process. p. 6

Multi-Agent Interaction

To mitigate redundant exploration behavior, a repulsion term is introduced into the reward function to discourage agents from selecting the same targets. The reward function is defined as the sum of the sampled observations z˜ci, minus the repulsion cost over all cells along the trajectory τj. p. 7 This method promotes slight detours toward highly rewarding cells along the trajectory. p. 7

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Decentralized terrain-aware Thompson sampling algorithm for multi-drone SAR applications enables agents to balance exploration of uncertain regions and exploitation of high-probability areas with a goal of locating a lost person.

  2. Object detection uncertainty based on fractional forest coverage allows the system to account for reduced detection performance in densely vegetated areas by incorporating observation uncertainty into the Bayesian updates.

  3. The framework uses terrain-informed priors derived from established SAR models (displacement ring, elevation, watershed, dispersion angle, and track offset) to assign informed priors for search of lost persons in large remote environments.

  4. The system employs a repulsion term in the reward function to discourage agents from selecting the same targets, thereby promoting spatial diversity and improving overall search efficiency by avoiding redundant exploration behavior.

  5. The uncertainty weighting mechanism slows the rate of convergence for cells with high uncertainty, ensuring they remain viable candidates for continued exploration, enabling agents to dynamically return to regions with high variance during search.

Related papers