Source-linked AI summary
SwarmNxt: Open-source Software-Hardware Platform for Fast and Agile Aerial Swarms
Charbel Toumieh, Niel Mistry, Benjamin Jarvis, Simon Jeger, Peize Liu, Shaojie Shen, Dario Floreano
TL;DR
Aerial swarm research needs robust autonomy and platforms that support onboard computation and scalable multi-robot software management. SwarmNxt combines OmniNxt hardware with deployment tooling and a ROS 2 autonomy stack, and its indoor experiments achieved collision-free collective flight while exposing current depth, GPU, and localization constraints.
Problem
Aerial swarm research is constrained by limited open platforms with sufficient onboard computation, sensing, and scalable tools for deploying and maintaining multiple autonomous drones.
Method
SwarmNxt combines OmniNxt hardware with hardware assembly guidance, Ansible-based fleet orchestration, and a ROS 2 stack integrating planning, control, and depth estimation.
Results
Six-drone and four-drone indoor experiments recorded no inter-agent or obstacle collisions during decentralized planning and collective flight.
Takeaways & Limitations
SwarmNxt provides an open and adaptable infrastructure for reliable, scalable physical aerial-swarm experimentation.
Takeaways & Limitations
The system relies on external motion capture, while depth estimation has limited resolution and near-exclusive GPU use.
Abstract
from arXiv · showhide
Aerial robot swarms have the potential to transform time-critical safety, security, and search-and-rescue operations. By coordinating multiple robots, they can rapidly survey disaster sites, map collapsed or GPS-denied environments, and search cluttered areas faster than a single robot, reducing response times and minimizing risks to first responders. Realizing this potential, however, requires robust autonomous swarm navigation, which remains an active research challenge. Progress is further constrained by existing platforms, as commercial drones are often closed-source or lack the onboard computational resources needed for agile, vision-based collective flight. Moreover, developing, deploying, and maintaining software across multiple aerial robots requires significant engineering effort. To address these challenges, we present SwarmNxt, an open-source software platform built on the open-source OmniNxt drone hardware. SwarmNxt provides an end-to-end toolkit, including detailed hardware assembly instructions with a video tutorial, automation tools for parallel software deployment and swarm-wide updates, and a ROS 2-based framework for autonomous navigation. The platform integrates state-of-the-art control, planning, and depth estimation into a single ROS 2 multi-agent system, providing an open research infrastructure for physical swarm experimentation. We validate SwarmNxt through two real-world experiments: a six-drone swarm performing decentralized planning with high-speed inter-drone collision avoidance, and a four-drone swarm executing collective flight with onboard depth estimation in an obstacle-filled environment. Both experiments were run indoors with global position from external motion capture; perception, planning, and control run onboard.
I. INTRODUCTION
SwarmNxt addresses the lack of open, research-grade platforms that combine onboard autonomy capabilities with scalable multi-drone deployment. Built on OmniNxt, it provides an open-source framework intended to streamline swarm experimentation.
- Swarm autonomy remains challenging because reliable navigation, onboard perception, and scalable multi-agent experimentation require capabilities that existing platforms and workflows do not consistently provide.
- Existing platforms trade off deployment simplicity against research capabilities, with commercial systems often closed-source and open alternatives lacking compute, sensing, or complete swarm tooling.
- SwarmNxt provides an open-source framework that streamlines the transition from hardware assembly to multi-robot deployment.
- The platform builds on OmniNxt hardware and combines hardware documentation, deployment automation, and an extensible ROS 2 autonomy framework.
II. FRAMEWORK ARCHITECTURE
The framework architecture includes a deployment pipeline whose stages are documented with an overview of the time required for swarm assembly and deployment.
- SwarmNxt organizes swarm deployment as a staged pipeline spanning the framework’s assembly and deployment process.
- Table II provides a temporal overview of the time required for each deployment stage.
A. Physical Assembly
SwarmNxt addresses the initial challenge of procuring and assembling multiple identical aerial agents with detailed hardware resources and guided instructions.
- Procuring and assembling multiple identical agents is identified as the initial technical hurdle in swarm research.
- The platform supplies a bill of materials, vendor links, atomic build instructions, and a video tutorial for OmniNxt assembly.
- Researchers without specialized mechanical expertise can assemble an OmniNxt drone in approximately five hours.
B. Swarm Software Orchestration
SwarmNxt uses Ansible-based automation to maintain software consistency and operational readiness across an expanding drone fleet.
- Software consistency and operational readiness become significant bottlenecks as the number of drones increases.
- Four work-specific Ansible playbooks perform parallel configuration and validation across the fleet.
- The individual setup playbook assigns unique hostnames and configures serial access and power profiles for each drone.
2) Parallel Software Deployment:
SwarmNxt automates synchronized software setup, updates, validation, and post-flight data collection across the drone fleet.
- 2) Parallel Software Deployment:: Parallel playbooks clone and build the latest ROS 2 autonomy packages across the swarm while automatically installing dependencies and updating only changed packages.Ansible idempotency allows the setup process to be rerun when packages change.
- 2) Parallel Software Deployment:: Pre-flight checks verify camera synchronization and focus, communication latency, and time synchronization before takeoff.
- 2) Parallel Software Deployment:: Post-flight automation retrieves logs and ROS-bag files simultaneously, centralizes them on the Host PC, and reports key flight-performance diagnostics.The diagnostics include MPC tracking error, autonomy-stack computation times, and minimum inter-agent distance.
C. ROS 2 Autonomy Stack
The ROS 2 autonomy stack combines isolated multi-agent communication with onboard planning, control, mapping, and depth estimation for decentralized swarm flight.
- C. ROS 2 Autonomy Stack: Each drone uses an isolated ROS Domain ID, while domain bridges selectively exchange essential information to scale communication and support decentralized collision avoidance.Planned trajectories are broadcast between agents, while planning, mapping, and control remain onboard.
- C. ROS 2 Autonomy Stack: HDSM performs mapping and motion planning, adaptive MPC tracks trajectories at 100 Hz, and S2M2 estimates depth from stereo images.Together, these components form the platform’s integrated autonomous navigation stack.
- C. ROS 2 Autonomy Stack: S2M2 converts rectified stereo-image pairs into dense depth maps using per-pixel disparity, with the 26.5M-parameter model operating at 256×160 resolution.The small variant was selected to balance depth accuracy and real-time onboard performance.
- C. ROS 2 Autonomy Stack: The mapping pipeline removes neighboring-drone points before integrating depth data and uses uncertainty-aware voxel classification to preserve real-time map stability.Voxels are classified as free below -1, occupied above 1, or unknown otherwise.
2) Safety Supervision:
SwarmNxt provides per-drone safety supervision and centralized operational monitoring to protect experiments and identify fleet or software problems.
- 2) Safety Supervision:: The Safety Node commands landing when a drone leaves the predefined safe zone or its EKF variance exceeds a critical threshold.Low-battery landings are handled separately by the flight controller.
- 2) Safety Supervision:: A unified dashboard issues global or individual flight commands and monitors battery voltage, communication latency, and ROS-node status.It is used with Foxglove for detailed visualization of drone poses and planned trajectories.
III. EXPERIMENTAL VALIDATION
SwarmNxt was validated in two indoor multi-drone flights assessing inter-agent collision avoidance and obstacle avoidance with onboard depth perception.
- III. EXPERIMENTAL VALIDATION: The experiments used six drones for free-space collision avoidance and four drones for obstacle avoidance with depth perception in an 8×8×4 m motion-capture arena.The vision experiment used four drones because the remaining units lacked fully calibrated fisheye cameras.
- III. EXPERIMENTAL VALIDATION: 0.075 m mean tracking error and safe inter-agent distances were maintained across 2-minute flights, while mapping and planning met the 100 ms worst-case budget.The planner safety radius was 0.45 m, covering the 0.301 m maximum tracking error.
- III. EXPERIMENTAL VALIDATION: 95% GPU and 54% CPU utilization accompanied onboard depth estimation, leaving CPU headroom for additional modules.Depth throughput was approximately 7 Hz and was measured as free-running rather than against a deadline.
- III. EXPERIMENTAL VALIDATION: Both experiments used identical controller, planner, and mapper parameters, with depth estimation disabled in the free-environment scenario.
B. Free Environment
A six-drone swarm tested decentralized inter-agent collision avoidance through position swaps and continuous roaming in an open environment. Across approximately two hours of cumulative flight, no collisions were recorded despite communication delay and 0.2% packet loss.
- The six-drone experiment began with a circle-to-circle position swap that forced drones to converge toward the center and created strong aerodynamic disturbances.A continuous roaming phase then assigned random goal positions along the circle.
- No collisions were recorded during approximately 2 hours of cumulative flight despite communication delay and 0.2% packet loss.The authors attribute this outcome to the planner’s robustness under these conditions.
C. Environment with Obstacles
A four-drone swarm evaluated onboard vision-based depth estimation and obstacle avoidance in an environment with static obstacles. During 30 minutes of cumulative flight, no collisions with obstacles or between agents were recorded, using conservative planning and mapping tuning.
- The obstacle experiment deployed four drones with onboard vision-based depth estimation and continuously updated random goals along a circular trajectory.The protocol matched the roaming procedure used in the free-environment experiment.
- No collisions with obstacles or between agents were recorded during 30 minutes of cumulative flight.The result followed conservative tuning of the planning and mapping modules to avoid collisions.
IV. CONCLUSION
SwarmNxt is an open-source, adaptable aerial-swarm research platform built on OmniNxt that streamlines deployment and supports scalable multi-drone operation. Its current scope is bounded by limited depth-estimation resolution, GPU exclusivity, and reliance on external motion capture, while enabling further research toward autonomous outdoor swarms.
- SwarmNxt streamlines the deployment lifecycle from hardware procurement and assembly through multi-drone operation and log analysis.Its software orchestration tools and autonomy stack can also be adapted with minor changes to drones using NVIDIA Orin and PX4.
- The platform’s current depth pipeline has limited resolution, requires near-exclusive onboard GPU access, and relies on external motion capture for agile swarm flight.These constraints limit small-object detection at distance, concurrent perception modules, and fully self-contained localization.
- SwarmNxt provides an open infrastructure for advancing robust depth estimation, lightweight relative localization, and multi-agent exploration toward fully autonomous outdoor drone swarms.The stated future directions include challenging capabilities beyond the current indoor motion-capture-based setup.