Session: RA4.1: Robotics & Automation-Specialized Navigation & Control

Paper Number: 187405

187405 - Atlas: A Cooperative Aerial System for Large-Scale Search and Delivery 

Large-scale aerial searches in disaster-impacted and unstable environments present difficulties arising from constrained flight endurance, limited computational capacities onboard, and the challenge of precisely identifying targets using top-down photographs. This study proposes ATLAS (Autonomous Team-oriented Localization and Search System), a cooperative aerial framework that effectively incorporates perception, geometric localization, adaptive multi-drone planning, and autonomous payload deployment, hence delivering an end-to-end solution for comprehensive search operations and targeted delivery.

ATLAS is designed around a team-oriented architecture in which multiple aerial agents operate covering mission grounds parallely. The system’s perception module integrates real-time vision with geometric systems to transform aerial detections of humans into real world spatial coordinates. Each agent runs a custom-trained deep learning model for top-view aerial human detection onboard, enabling autonomous operation without external infrastructure or communication. Detected image-space coordinates are transformed into real-world positions using camera intrinsics and onboard pose estimation from UAV’s parameters. Rotation-based transformation matrix maps pixel locations into metric coordinates, which are then fused with global positioning data of UAV to compute precise geotagged locations of detected targets. The perception-to-navigation pipeline is connected through a disk-backed storage, allowing UAVs to transition from detection to required action.

To efficiently explore extensive regions, ATLAS introduces a strip-stealing path-planning strategy based on strip partitioning and dynamic workload redistribution. The search space is distributed into parallel strips initially divided and assigned equally across all agents. Unlike static partitioning, ATLAS enables continuous adaptation through a dual-layer behavioral model. Each agent operates under a stage reflecting remaining workload in terms of strips Donor or Acceptor , where Donor is yet to complete its initially assigned strips and Acceptor has completed its initial mission strips and is ready to accept the donated strips from another agent. Each agent also has a state reflecting remaining delivery capacity Executor or Scout, where executor has parcels it can deliver and scout has no parcels remaining. The combination of these dimensions governs agent behavior throughout the mission.

This strip-sharing mechanism minimizes overall mission time and prevents idle behavior, ensuring that all agents remain productive until full coverage is achieved. By modularising exploration responsibility from delivery capability, ATLAS supports uninterrupted search even when limited resource constraints arise. The resulting system dynamically redistributes exploration and execution, enabling adaptive and efficient coordination across all agents in each mission phase.

The targeted delivery is performed using a sequential payload release mechanism integrated into each drone. The mechanism is designed to support multiple in-mission payload releases. Payload release is triggered only after a human has been detected and their GPS location is obtained. This design enables a precise release of payload by hovering on the target. The payload dropping assembly consists of two servo operated mechanisms- a grabbing and a flap mechanism each holding a single payload box. Additional boxes are stacked above, enabling sequential release. 

Real-world flight experiments and comprehensive simulation using ArduPilot SITL were used to validate ATLAS. To verify the accuracy of perception, geometric localisation, autonomous navigation, and payload deployment pipelines, single-agent trials have been carried out. Stable collaboration with dynamic strip division and role-based coordination has been demonstrated in multi-drone experiments. While maintaining reliable target positioning and delivery, ATLAS substantially reduced mission completion time and lowered idle state as compared to single-drone operations and static multi-drone partitioning. The system demonstrated endurance and scalability for time-sensitive large-area search and targeted delivery missions by attaining fully autonomous end-to-end operation across detection, localisation, exploration, and payload release.

Presenting Author: Harshal Kolhe Visvesvaraya National Instiute of Technology, Nagpur

Presenting Author Biography: Harshal Kolhe is a prefinal year B.Tech student at VNIT Nagpur studying Chemical Engineering with a minor in IoT. He has a keen interest in aerial robotics, UAVs, and autonomous navigation. In addition to developing systems for agricultural monitoring and spraying, multi-agent disaster rescue, and GPS-denied quadcopter navigation using vision, optical flow, and range sensing, he has constructed several custom drones. He conducted research at IvLabs, VNIT (fixed-wing UAV design and autonomous flight) and IIT Roorkee (reinforcement learning for quadcopters with suspended payloads). Harshal is currently Vice Chairman of IvLabs and co-author of an IEEE BigData 2025 paper on LLM-based UAV mission generation.