The Swarm project has a clear multi-stage roadmap that outlines its evolution from basic simulations to real-world drone swarms. Each stage introduces new technical challenges for miners and moves the network closer to practical drone autonomy. The stages are summarized below:

Stage 0 – Baseline (“Boot-Up”): Goal: Deliver a minimal viable product where anyone can easily run a miner or validator. This initial phase focuses on fun and fairness – providing a reference implementation of the drone flight task that works out-of-the-box. The default code at this stage generates relatively simple random maps, and miners compete to improve flight routes based on the reward criteria (success, time, energy). Stage 0 establishes the core framework and ensures that even a naïve strategy can participate, seeding the network with activity.

Stage 1 – Static Map Difficulty Ramp-Up: Goal: Stress-test miners on harder static environments. In this phase, the flight challenges become tougher but remain in a static world (no moving obstacles yet). Key upgrades include a larger search space – the target pole can be placed further away, there may be elevation changes between start and goal, and more occluding obstacles are introduced. This forces miners to use more sophisticated planning to navigate around obstacles and handle longer distances. Stage 1 essentially dials up the difficulty gradually to push miners beyond the trivial solutions.

Stage 2 – Dynamic Obstacles & Moving Goals: Goal: Introduce temporal planning aspects by making the environment dynamic. In this stage, the simulator will add moving elements: for example NPC drones or birds that fly around as moving obstacles, and even a moving goal (the target platform might move along a path). Additional mechanics like collision penalties and designated no-fly zones come into play. Miners now must anticipate changes over time – their flight plans need to account for objects that won’t stay put. Stage 2 significantly increases complexity, as miners can no longer assume a static world and must plan trajectories in both space and time.

Stage 3 – High-Fidelity Simulation (“Simulation Jump”): Goal: Narrow the gap between simulation and reality. In Stage 3, the team plans to migrate from the current PyBullet simulator to a higher-fidelity simulation platform. They will explore options such as Gazebo, AirSim, Flightmare, or Pegasus – simulators that provide more realistic physics, sensor models, and possibly 3D environments closer to the real world. This “Sim Jump” will make the virtual drone flights more accurately reflect real drone behavior (including aerodynamics, turbulence, etc.). The purpose is to ensure that strategies developed in simulation remain valid when eventually transferred to physical drones. Upgrading the simulation may also involve more complex world rendering and environment physics, preparing the network for Stage 4.

Stage 4 – Generic Drone Flight (“Random Drone”): Goal: Enable miners to control a variety of drone models rather than one fixed drone type. By Stage 4, the plan is to introduce a fleet of different drone profiles into the tasks – each with unique characteristics like weight, motor power, battery capacity, and aerodynamics. Miners won’t know in advance which drone model they’ll get for a given task, so their algorithms must become adaptable or robust to different vehicle dynamics. This stage broadens the challenge: instead of optimizing for a single drone, miners must demonstrate generalizable flying skills. The addition of multiple real-world drone models in simulation paves the way for engaging with actual drone manufacturers and industry. In fact, the roadmap notes that around this stage, the team will start discussions with businesses to adapt the subnet’s output to real-world needs. The idea is to align what Swarm produces (e.g. autopilot policies) with what real drone companies or researchers can use.

Stage 5 – Long-Range Navigation (“Travel Missions”): Goal: Test drones on extended, GPS-denied missions that require endurance and multi-stage navigation. Stage 5 will introduce very large procedurally generated maps spanning multiple kilometers. Drones might have to fly through a sequence of connected map areas, essentially requiring stitching together multiple waypoints or sub-missions. Mid-flight battery management becomes crucial: the roadmap mentions adding mid-point recharging pads, so a drone could land and recharge en route to cover longer distances. Also, realistic sensor limitations will be simulated – e.g. magnetometer and barometer noise, meaning the drone’s onboard sensing of direction and altitude might be imprecise. Moreover, “GPS-denied” implies the drone may have to navigate by dead-reckoning or onboard sensors only (simulating environments like indoors or areas without GPS). All these additions make Stage 5 a test of both algorithmic endurance and reliability over long durations.

Stage 6 – Interceptor Missions: Goal: Evolve beyond passive navigation to an active pursuit scenario. In Stage 6, miners will face tasks where their drone must intercept another moving drone within a time limit. This is essentially a high-speed chase challenge. The target drone will follow some scripted path, and the miner’s drone (the pursuer) has to catch up and intersect that path in mid-air. The reward at this stage will be tuned to prioritize capture speed (how quickly the target is caught) and safety (no collisions or reckless maneuvers). This scenario pushes miners to develop interception and targeting strategies (potentially involving predictive modeling of the target’s trajectory). It’s a step towards defense or security applications (like intercepting rogue drones), adding a competitive agent to the environment.

Stage 7 – Controlled Real-World Pilots: Goal: Validate Swarm’s AI pilots on real hardware in controlled settings. By Stage 7, the best algorithms from the simulation will be tested on actual drones in a safe, controlled environment (such as an indoor drone test lab). The team plans to use an indoor motion-capture arena (which provides precise ground-truth position tracking) and run drones with a Linux-based autopilot stack to deploy the policies learned in simulation. This stage will likely involve a feedback loop: after real flight tests, the flight logs can be automatically uploaded and notarized on-chain as proof of real-world performance. The idea is to ensure the on-chain rewards truly correspond to algorithms that work not just in simulation but in reality. Stage 7 is essentially a research validation phase (“Track 6-R” in the roadmap) to bridge the sim-to-real transfer – a critical step before going fully public or commercial.

Stage 8 – Commercial Partnerships & Services: Goal: Turn the open-source drone pilot into real-world value. In the final envisioned stage, Swarm will reach out to industry and form partnerships to deploy its technology in practical applications. Potential domains mentioned include: last-mile logistics (delivery drones), industrial inspection and mapping (autonomous survey drones), public safety and emergency response (search-and-rescue UAVs), defense and counter-UAS (military or security drone swarms), UAV manufacturers & autopilot vendors (integrating Swarm’s algorithms into commercial drone platforms), and academia/certification bodies (for research and setting standards). Essentially, the best performing open-source autopilot solutions coming out of Swarm could be offered as services or be licensed/adapted for these sectors. The focus shifts to sustainability – ensuring the project can generate value (and possibly funding or revenue streams) that keep it running long-term. Stage 8 marks the maturation from a purely research project into an applied technology platform for aerial autonomy.