How AI Pilots Flying Cell Towers to Better Coverage

Mobile base stations — drones or ground vehicles equipped with cellular transmitters — are the future of flexible, on-demand wireless networks. They can hover above disaster zones, follow crowds at concerts, or boost coverage in rural areas. But getting them to fly the right path while optimizing signal strength is a nightmare of physics, interference, and safety. Now, a team from the University of Oulu has built an AI system that doesn’t just react — it imagines the future.

In simulations, their framework, called SafeDreamer, improved average signal quality by 7–9% over standard reinforcement learning methods. That may sound small, but in wireless networks, a 5% gain is often considered transformative. More importantly, it did so while respecting no-fly zones, power limits, and mission constraints — all learned through trial, error, and internal simulation.

This isn’t just about better phone service. It’s about building intelligent infrastructure that can plan ahead, avoid danger, and adapt in real time — a step toward autonomous systems that don’t just act, but think.

The Science

The challenge begins with a simple question: where should a flying base station go, and how should it shape its signal, to give the best possible coverage over time?

The answer is deceptively complex. Every movement changes the wireless environment. As the base station shifts position, the signal strength to each user fluctuates. Interference between users shifts. The optimal beamforming — the way the antenna focuses its energy — must adapt instantly. And because the system is dynamic, today’s decision affects tomorrow’s options. Move too far north now, and you might not be able to reach a user in the south later. Enter a restricted zone, and the mission fails.

The researchers, led by Abdolrasoul Sakhaei Gharagezlou and colleagues at the Centre for Wireless Communications in Finland, framed this as a constrained Markov decision process — a mathematical model for sequential decision-making under uncertainty. The goal: maximize the time-averaged sum signal-to-interference-plus-noise ratio (SINR) across all users, subject to:

• Mobility constraints (maximum speed)
• Operating region boundaries
• A circular restricted zone (RZ) to avoid
• Initial and final position requirements
• Transmit power limits

SINR is the gold standard in wireless: it measures how strong a user’s signal is relative to noise and interference. Higher SINR means faster data, fewer dropped calls, better video streaming. Maximizing it over time ensures sustained performance, not just a momentary peak.

Solving this with traditional optimization is nearly impossible. The problem is non-convex, stochastic, and high-dimensional. Model-free reinforcement learning (MFRL) methods like PPO or DQN can learn policies through trial and error, but they require massive amounts of real-world interaction — impractical for drones that can crash or violate airspace.

So the team turned to world models — AI systems that learn an internal simulation of the environment. Specifically, they built a variant of Dreamer, a deep reinforcement learning agent that learns a compact latent representation of the world and uses it to plan by “imagining” future trajectories.

Their version, SafeDreamer, adds two key innovations: feasibility-preserving action mappings and safety-cost signals. These allow the agent to learn not just performance, but safety — without violating constraints during training.

What They Found

The simulation setup involved a single mobile base station (MBS) with a 4-element antenna array serving 3 ground users over 100 time slots. The MBS had to travel from a start to end point, avoid a central restricted zone, and maximize SINR under realistic channel dynamics, including time-varying fading and distance-dependent path loss.

The agent’s observations included its own position, user locations, distances, channel gains, past SINR, and relative vectors to the final destination and restricted zone. Actions were composed of 2D displacement and per-user power allocation, from which beamforming vectors were derived using regularized zero-forcing — a standard technique to suppress interference.

The world model — built around a Recurrent State-Space Model (RSSM) — learned to encode observations into latent states, predict future rewards and safety costs, and reconstruct observations. Training alternated between collecting real trajectories and optimizing the policy using imagined rollouts from the learned world model.

The results were clear. Over multiple runs, SafeDreamer outperformed two strong baselines:

• Proximal Policy Optimization (PPO): a popular on-policy RL algorithm
• Deep Q-Network (DQN): a foundational off-policy value-based method

Average Sum SINR Improvement Over Baselines

Label	Value
SafeDreamer	1.09
PPO	1.02
DQN	1

SafeDreamer achieves 9% higher average sum SINR than DQN and 7% higher than PPO in simulation.

The key metric was time-averaged sum SINR. SafeDreamer achieved an average improvement of 7% over PPO and 9% over DQN. This gain came not from brute-force exploration, but from smarter, longer-horizon planning. By imagining dozens of future steps, the agent could anticipate how a current move would affect future signal conditions — something model-free methods struggle with.

Even more striking was the improvement in sample efficiency. SafeDreamer reached high performance with far fewer environment interactions. In real-world terms, this means fewer drone flights, less energy, and faster deployment.

Safety Cost Over Training Episodes

Label	Value
Episode 1	0.45
Episode 10	0.32
Episode 20	0.21
Episode 30	0.15
Episode 40	0.11
Episode 50	0.08

Safety cost decreases steadily during training, showing that SafeDreamer learns to avoid constraint violations over time.

But performance means nothing if the system breaks the rules. Here, SafeDreamer shone. The safety-cost learning framework — which penalized violations of operating boundaries, restricted zones, and final-position constraints — ensured that the agent learned to avoid danger during training, not just at test time.

The total safety cost — a normalized sum of constraint violations — was significantly lower than both baselines. PPO and DQN occasionally drifted into restricted zones or failed to reach the final destination; SafeDreamer did not.

shows example trajectories. The red circle is the restricted zone. SafeDreamer (green) smoothly navigates around it, while PPO (blue) cuts too close, and DQN (red) briefly enters the zone. The final position is consistently reached only by SafeDreamer.

This reliability isn’t accidental. The safety cost function included four components:

1. Operating-region violation cost: penalizes leaving the allowed area
2. Final-position reachability cost: ensures the destination is reachable within remaining time
3. RZ violation cost: penalizes entering the restricted zone
4. Final-position error cost: measures deviation from the target at mission end

These were combined into a single safety return, and the policy was optimized to maximize communication performance subject to a budget on expected safety cost. This is the essence of constrained reinforcement learning — trading off reward against risk.

Why This Changes Things

At first glance, this is a technical advance in wireless optimization. But zoom out, and it’s a blueprint for a new kind of intelligent infrastructure.

Today’s networks are static. Cell towers are fixed. Traffic is managed by pre-programmed rules. When demand spikes, congestion follows. When a tower fails, coverage drops. Mobile base stations promise flexibility — but only if we can control them intelligently.

SafeDreamer represents a shift from reactive to proactive control. It doesn’t just respond to current signal conditions; it anticipates how its actions will shape the future. It plans. It imagines. And it does so while respecting hard constraints — the kind that separate research from reality.

Consider disaster response. A fleet of drones is deployed to restore communication after an earthquake. Roads are blocked, power is out, and every second counts. A model-free agent might maximize immediate coverage but crash into a building or run out of battery before reaching survivors. SafeDreamer, by contrast, would weigh the long-term consequences of each move, balance signal quality with safety, and adapt as conditions change — all learned from simulated experience.

This is the power of world models. Unlike traditional RL, which learns a policy directly from rewards, world models learn a dynamics model — a kind of internal physics engine. Once trained, the agent can simulate thousands of futures in seconds, refining its strategy without real-world risk.

The implications extend beyond drones. Autonomous vehicles, robotic warehouses, smart grids — all face similar trade-offs between performance and safety. SafeDreamer’s architecture, with its dual focus on reward and cost, offers a template for building systems that are not just smart, but responsible.

It also addresses a critical bottleneck: sample efficiency. In the real world, every interaction has a cost — fuel, time, wear and tear. SafeDreamer’s ability to learn from imagined trajectories means it can achieve high performance with minimal real-world data. For industries where trial and error is expensive or dangerous, this is transformative.

Compare this to current 5G and 6G research. Many proposals rely on centralized optimization or heuristic rules. They work in simulations but falter in the real world, where conditions change too fast. SafeDreamer, by learning from experience and planning ahead, is built for uncertainty.

And unlike black-box AI systems that violate constraints unpredictably, SafeDreamer learns to respect them. The safety-cost signal isn’t a patch — it’s baked into the learning process. This is crucial for regulatory approval and public trust.

What’s Next

The study is a simulation, and real-world deployment brings new challenges. Wireless channels are noisier, drones have mechanical limits, and GPS can fail. The team used a simplified Gauss-Markov fading model; real environments have multipath, shadowing, and human movement.

Still, the framework is general. It could be extended to multi-agent settings — a swarm of drones coordinating coverage. The safety-cost approach could incorporate new constraints: battery life, weather, air traffic.

One limitation is the reliance on a fixed mission duration and known user locations. In practice, users move, and missions may be open-ended. Future work could integrate user mobility prediction or online learning.

Another open question is scalability. The current setup serves 3 users. Real networks have dozens or hundreds. While beamforming can scale, the action and observation spaces grow with user count. Techniques like attention or graph neural networks could help.

The use of regularized zero-forcing beamforming is practical but suboptimal in highly dynamic environments. End-to-end learning of beamforming directions — rather than relying on a fixed algorithm — could unlock further gains.

Perhaps the biggest challenge is trust. How do we verify that an AI system won’t violate safety constraints in a novel situation? The paper uses a safety budget, but real-world systems may need formal verification or explainability tools.

Still, the direction is clear. We’re moving from static networks to cognitive ones — systems that perceive, plan, and adapt. SafeDreamer is a step toward that future: not just a smarter drone, but a more thoughtful one.

As wireless networks become more dynamic, the algorithms that control them must become more intelligent. This paper shows that the future of connectivity isn’t just faster data — it’s smarter decisions.