PaperlayerView exact revision on arXiv

Source-linked AI summary

D3ARC: Time-Critical Distributed Disaster Detection for Asynchronous Cooperative Multi-Robot Systems

Nikolaos Koursioumpas, Lina Magoula, Nancy Alonistioti, Ramin Khalili

arXiv:2609.07350v1cs.ROcs.AIcs.CVcs.MA

TL;DR

Wildfire monitoring is constrained by coverage, cost, personnel risk, and time, while operational delays can reduce the opportunity for early detection. D3ARC combines asynchronous controller–robot decisions with cooperative mechanisms and forward-looking strategy evaluation. The paper reports mission success up to 94% with 89.4% detection confidence.

  • Problem

    Traditional wildfire monitoring is constrained by limited spatial coverage, high operational costs, personnel risk, and time-consuming operations that can reduce early-detection opportunity.

    Source: §I · p. 1

  • Method

    D3ARC combines asynchronous controller–robot decisions with safety, coverage, cooperation, early stopping, and distributed evaluation of candidate strategies.

    Source: §I · pp. 1–2

  • Results

    Up to 94% mission success with 89.4% detection confidence; multi-robot cooperation improves coverage speed and mission duration and enables earlier, reliable detections.

    Source: §VII · p. 12

  • Takeaways & Limitations

    Multi-robot cooperation improves coverage speed and mission duration and enables earlier, reliable wildfire detections.

    Source: §VII · p. 12

Abstract

from arXiv · show

Climate change is increasing the severity and unpredictability of natural disasters. In time-critical crises such as wildfires, traditional monitoring practices remain limited by coverage, cost, and personnel risk, paving the way for autonomous and adaptive monitoring solutions. Within this context, this paper introduces D3ARC, an asynchronous distributed hierarchical framework for time-aware and reliable wildfire detection. D3ARC integrates multiple robotic agents that cooperate under uncertainty through distributed perception, shared situational awareness and coordinated actions. A remote controller asynchronously decides upon each robot's motion, while each robotic agent senses the environment and decides where and how to execute the wildfire detection. All robotic operations require time, and as time progresses, wildfires continue to spread, reducing the opportunity for early intervention. As such, all agents share a common objective: to detect a wildfire with a certain performance threshold as fast as possible and within a time limit. D3ARC integrates mechanisms for safe navigation, coverage efficiency, cooperation and reliability. It introduces a forward-looking capability that allows agents to anticipate the future by evaluating candidate strategies before execution. The framework is evaluated through realistic robotics simulations, ablation studies, and baseline comparisons, achieving an overall mission success up to 94% with 89.4% detection confidence.

I. INTRODUCTION

D3ARC addresses wildfire monitoring’s coverage, cost, personnel-risk, and time constraints with asynchronous cooperative robotics for fast, reliable detection. It combines forward-looking decisions and operational mechanisms; the related-work discussion identifies gaps in unified treatment of time, reliability, network variability, and resources.

I. INTRODUCTION

  • Robotic operations consume time across sensing, computation, communication, and movement, while accumulated delays can reduce early-detection opportunity as wildfires spread.The introduction cites a 10-minute detection target after ignition.Source: §I · p. 1
  • D3ARC combines an asynchronous controller and robot decisions for wildfire detection, with robots choosing where and how to execute detection.The controller assigns motion, while agents select between onboard and remote execution and choose a detection model.Source: §I · pp. 1–2
  • The framework integrates obstacle avoidance, coverage efficiency, call-for-aid cooperation, adaptive early stopping, and penalties for infeasible strategies.Coverage scoring redirects robots from repeatedly visited regions toward promising or unexplored areas; early stopping avoids unnecessary motion and time cost.Source: §I · p. 2
  • Distributed neural regression models evaluate candidate strategies before execution so agents can estimate shared outcomes and trade off detection confidence against mission duration.The remote controller also anticipates robots’ subsequent actions when estimating outcomes.Source: §I · p. 2
  • D3ARC addresses heterogeneous perception capabilities, and the introduction reports cooperation and robust performance under unseen scenarios and setups.The reported performance dimensions include detection confidence, mission duration, and overall mission success.Source: §I · p. 2

II. RELATED WORK

  • The related-work comparison identifies limited asynchronous cooperation and few works using neural-network forward-looking reasoning, with no unified framework covering time, reliability, network variability, and resource availability.These gaps motivate D3ARC’s distributed hierarchical cooperative decision-making.Source: §II · pp. 2–3, Table I · p. 3

III. SYSTEM MODEL

The system model assigns complementary roles to a remote controller and UAV robots, whose sensing, computation, and communication capabilities affect wildfire detection. Time advances asynchronously as robots finish tasks, and the controller can assign motion without waiting for the entire fleet.

III. SYSTEM MODEL

  • UAV robots sense the environment and can detect wildfires onboard, while the remote controller asynchronously handles robot motion and can assist with remote detection.Robots may offload sensing data to the controller for its wildfire detection algorithm.Source: §III · p. 3
  • A new timestep begins when one or more robots complete tasks; robots ready within a tolerance are grouped, and the controller assigns each ready robot its next motion command without waiting for others.The command’s execution time varies across robots and timesteps.Source: §III · p. 3
  • Robots sense for a fixed period and have heterogeneous detection capabilities, which external factors such as blur, illumination, reflections, and viewpoint artifacts can limit.Robots are classified as standard-detection or limited-detection according to their ability to produce reliable detections.Source: §III · p. 3
  • Available AI algorithms differ in computational complexity, average performance, and processing time, while robot computational capacity and CPU resources are represented explicitly.Algorithm complexity is measured in FLOPs; the model also represents FLOPs per cycle and effective switched capacitance.Source: §III · p. 3
  • Robots may offload sensing data over a time-varying communication channel whose model includes noise, channel gain, bandwidth, transmission power, and achievable data rate.Channel fading is assumed constant during each timestep.Source: §III · p. 3

IV. PROBLEM FORMULATION

The formulation seeks cooperative wildfire detection at a required confidence as quickly as possible within a time limit. It models a sequence of interdependent remote-controller and robot decisions, including movement, sensing, and whether detection runs locally or remotely.

IV. PROBLEM FORMULATION

  • The objective is to reach the required wildfire-detection confidence as fast as possible under an imposed time limit, not merely to satisfy the confidence threshold.The controller and robots make hierarchical, distributed, interdependent decisions with complementary parameters and different views of the environment.Source: §IV · p. 3
  • The formulation treats RC and robot decisions as coordinated because each agent controls complementary parameters and acts with a different view of the same environment.These interdependent choices form the chain needed to meet the confidence and time objective.Source: §IV · p. 3

A. Requirements for Robot Movement and Sensing

  • Robot motion includes hovering, horizontal, and vertical movement, each with distinct time costs.The model defines movement time, distance traveled, and robot speed for each movement type.Source: A. Requirements for Robot Movement and… · p. 4
  • The decision chain begins with the remote controller’s asynchronous motion command, followed by each robot’s choice to offload detection or run it locally and select an AI model.The robot makes its execution choice after receiving the motion command and before sensing.Source: §IV · p. 3, A. Requirements for Robot Movement and… · p. 4
  • Figure 1 depicts the stages and dependencies of the controller–robot decision sequence and when each decision is made.The controller’s motion command precedes the robot’s offloading and model-selection decisions.Source: Figure 1 · p. 4, A. Requirements for Robot Movement and… · p. 4
  • After completing its motion command, each robot senses the environment for a fixed time period using its available sensors.The model assumes this sensing period is sufficient to acquire newly generated sensing data.Source: A. Requirements for Robot Movement and… · p. 4

B. Requirements for Local Execution

The section models local and remote detection time, then defines a temporally stable confidence score and its cooperative aggregation across asynchronous robots.

B. Requirements for Local Execution

  • Local execution time is modeled for processing sensed samples with a selected algorithm, assuming perfect parallelization across the robot’s CPU cores.The sample count is the number collected during the sensing period.Source: B. Requirements for Local Execution · p. 4

C. Requirements for Offloading

  • Offloading time accounts for transmitting sensing data to the remote controller, whose processing time is assumed negligible because of its high computational capacity.The achievable data rate is modeled in the Appendix.Source: C. Requirements for Offloading · p. 4

D. Detection Confidence

  • The Stable Confidence Score (SCS) captures detection stability over sensing window s using the selected local or remote confidence and its variation.Individual scores are combined into cooperative confidence; the three-robot timeline illustrates that different robots contribute at different asynchronous timesteps.Source: D. Detection Confidence · pp. 4–5, Figure 2 · p. 5

E. Problem Objective

The objective minimizes mission time until cooperative detection confidence reaches its threshold, subject to a time limit and constraints on performance and motion.

E. Problem Objective

  • The objective minimizes total mission time until cooperative confidence first reaches the required threshold, subject to the imposed time limit.The first threshold-reaching timestep defines the stopping point.Source: E. Problem Objective · p. 5
  • The formulation constrains cooperative detection performance, robot motion commands to hardware bounds, and elapsed wall-clock time to the imposed threshold.The selected motion commands and robot decisions are the decisions made during each timestep.Source: E. Problem Objective · p. 5
  • Penalties address violations of minimum detection confidence, motion bounds, and timestep duration, and the complete objective incorporates those penalties.Objective and penalty terms are min-max normalized and divided by their means before weights are applied.Source: E. Problem Objective · p. 5

V. D3ARC: DISTRIBUTED HIERARCHICAL INTELLIGENCE FOR COOPERATIVE ROBOTS

D3ARC uses an offline warmup followed by online execution in which a remote controller and robots make interdependent decisions from different views of the environment.

V. D3ARC: DISTRIBUTED HIERARCHICAL INTELLIGENCE FOR COOPERATIVE ROBOTS

  • D3ARC has offline warmup and online real-time execution, with the controller and robots’ different views and actions evaluated jointly through a reward function.The controller has a higher-level environmental perspective; robots have interaction-driven views.Source: §V · p. 5, A. State, Action, Reward · p. 5

A. State, Action, Reward

  • At each timestep, the controller’s state includes robot position and pose, distance, cooperative confidence, and elapsed time, while robot state also includes computing and communication resources.Robot state includes CPU frequency ratio, allocated bandwidth, and achievable data rate.Source: A. State, Action, Reward · pp. 5–6
  • The controller selects robot movement commands, while each robot chooses whether to offload detection and which model to use; all agents receive a mutual reward.The reward promotes cooperative detection confidence toward its threshold while penalizing action-related constraint violations.Source: A. State, Action, Reward · p. 6

B. Offline Warmup Phase

D3ARC trains forward-looking reward estimators offline, then uses them for asynchronous real-time decisions. Its online mechanisms coordinate sensing and coverage while supporting safe motion and adaptive mission completion.

B. Offline Warmup Phase

  • Offline warmup trains separate supervised RFNNs for the remote controller and robots to estimate shared reward from their respective states and actions.Collected rewards provide ground truth; the pretrained models are used for online execution.Source: B. Offline Warmup Phase · p. 6

C. Online Real-Time Execution Phase

  • At each timestep, the controller evaluates sampled candidate actions and robots evaluate all possible actions, greedily selecting those with the highest estimated reward.The controller samples N alternatives from the feasible action space; robots evaluate all possible actions.Source: C. Online Real-Time Execution Phase · p. 6
  • A robot-side 360° obstacle-avoidance mechanism halts motion toward obstacles detected within the minimum safety distance.The robot uses its local sensing perspective rather than blindly following a controller command.Source: C. Online Real-Time Execution Phase · p. 6
  • Coverage prioritization penalizes repeated visits to uninformative regions while assigning lower penalties to promising regions and intermediate penalties to unexplored ones.The penalty is based on region history, including visits and observed detection performance.Source: C. Online Real-Time Execution Phase · p. 7
  • Call-for-Aid redirects nearby robots toward a high-confidence but sub-threshold observation so they can collect complementary views, while distant robots continue their current operations.Participating robots align their movements with the reported position and orientation, then return to normal operation after reaching the region.Source: C. Online Real-Time Execution Phase · p. 7
  • Adaptive early completion makes reliable detections trigger hovering, ends the mission after confidence persists for the required timesteps, and adjusts action ranges for intermediate confidence.Persistent intermediate confidence first narrows exploration around the current viewpoint, then expands the action range if it continues.Source: C. Online Real-Time Execution Phase · pp. 7–8

VI. PERFORMANCE EVALUATION

The evaluation measures whether cooperative UAV missions detect a wildfire at the required confidence as quickly as possible and within a time limit. It examines time–reliability trade-offs, cooperation, and comparisons with non-cooperative and heuristic baselines.

VI. PERFORMANCE EVALUATION

  • Missions use up to four UAVs and a remote controller, with the shared objective of meeting a required detection confidence quickly and within a predetermined time limit.UAVs carry optical cameras, LiDAR, GPS, an IMU, and two computer-vision models; the controller also has a computer-vision model.Source: §VI · p. 8
  • The evaluation focuses on time versus mission reliability, hierarchical coordination and cooperation, and comparison with non-cooperative and heuristic baselines.Source: §VI · p. 8
  • Table III summarizes the evaluated metrics, which are averaged across missions and robots.The look-forward accuracy metric is introduced as part of the evaluation.Source: §VI · p. 8, Table III · p. 8

A. Simulation Setup

D3ARC is evaluated in a physics-aware Gazebo/ROS2 wildfire simulation with variable resources, locations, and detection capabilities. The configuration section reports offline RFNN training and fixed online parameters.

A. Simulation Setup

  • The Gazebo 11 and ROS2 simulation uses rigid-body dynamics and collision physics in a 530 x 530 x 165 m mountainous forest scene containing a 46 m wildfire with smoke.The scene was designed in Blender.Source: A. Simulation Setup · p. 8, Figure 3 · p. 8
  • UAVs have two custom wildfire-detection CNNs with different computational complexity, while the remote controller uses only the Full CNN.Both models are available to UAVs.Source: A. Simulation Setup · pp. 8–9
  • The wireless model links each robot to a remote controller collocated with a single base station and accounts for distance-dependent loss, shadow fading, bounded bandwidth, and spectral efficiency.Source: A. Simulation Setup · p. 9
  • Missions vary resource availability, wildfire locations, and detection capabilities; up to half the robots may have detection confidence capped below the mission threshold.The simulation includes 14 possible wildfire locations, one incident per mission, and two possible base-station locations.Source: A. Simulation Setup · p. 9

B. D3ARC Configuration Setup

  • Offline RFNN training uses 115K samples, with 85%, 10%, and 5% assigned to training, validation, and evaluation; tuned parameter values remain fixed across online missions.Training uses Adam at learning rate 5e-06 for up to 2000 epochs, with dropout and early stopping.Source: B. D3ARC Configuration Setup · p. 9, Table IV · p. 9, Table V · p. 9

C. Time-Aware Multi-Robot: Evaluation Results

Under a three-minute threshold, cooperative D3ARC outperforms the CH-RBS heuristic baseline on mission success and completion time. Across robot counts, cooperation improves mission success and other time- and detection-related measures; single-robot results remain limited by timely access to promising regions.

Results 1: Single Robot Time Limit Sensitivity

Results 2: Impact of Multi-Robot Cooperation under a Strict Time Limit

Results 3: Baseline Comparison

VII. CONCLUSIONS

D3ARC coordinates robots and a remote controller for time-aware wildfire detection, with a forward-looking approach and mechanisms supporting safe, efficient cooperation. Multi-robot experiments report mission success up to 94% with 89.4% detection confidence.

VII. CONCLUSIONS

  • D3ARC coordinates multiple robots with a remote controller: the controller decides each robot’s motion, while robots decide where and how to perform detection.Agents aim to detect wildfire at a target performance as quickly as possible within an imposed time limit.Source: §VII · pp. 11–12
  • A forward-looking approach and four mechanisms support safe operation while promoting exploration, cooperation, and system efficiency.Source: §VII · p. 12
  • Mission success reaches 94% with 89.4% detection confidence; comparisons with D3AR and CH-RBS confirm the merits of D3ARC’s adaptive, confidence-driven, time-aware decision-making.Multi-robot cooperation improves coverage speed and mission duration and enables earlier, reliable detections.Source: §VII · p. 12

APPENDIX

The appendix documents the achievable-data-rate model, UAV, computer-vision, and network configurations, and the H-RBS decision rule.

A. Achievable Data Rate Modeling

  • The achievable-data-rate model describes path loss using reference distance and carrier frequency, and defines robot–controller distance and the path-loss exponent.Source: A. Achievable Data Rate Modeling · p. 13

B. UAV Robot Specifications

C. Computer Vision Detection Model Configuration

D. Wireless Communication Network Specifications

E. Heuristic Rule-based Strategy (H-RBS)

  • H-RBS uses a random walk and compares local-execution and offloading ratios against thresholds 0.85 and 0.65 to select a strategy.If both ratios exceed their thresholds, H-RBS chooses randomly; if neither does, it selects the strategy whose ratio is closest to its threshold.Source: E. Heuristic Rule-based Strategy (H-RBS) · p. 13
Loading 2609.07350v1…