⚡ electrical engineering

Flying by Inference: Active Inference World Models for Adaptive UAV Swarms

This paper proposes an expert-guided active inference framework that transforms multi-UAV trajectory planning from a combinatorial optimization problem into a hierarchical probabilistic inference task, enabling adaptive swarms to generate smooth, collision-aware behaviors by learning from expert demonstrations and minimizing deviation from reference distributions without re-running offline optimizers.

Original authors: Kaleem Arshid, Ali Krayani, Lucio Marcenaro, David Martin Gomez, Carlo Regazzoni

Published 2026-05-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kaleem Arshid, Ali Krayani, Lucio Marcenaro, David Martin Gomez, Carlo Regazzoni

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the director of a massive, high-speed delivery service with hundreds of drones. Your job is to tell them where to go, in what order, and how to fly without crashing into each other or buildings. Usually, doing this math for hundreds of drones at once is like trying to solve a giant, impossible puzzle every single time a new package arrives. If you wait to solve the whole puzzle, the drones sit idle. If you try to solve it instantly, you might make a mistake.

This paper proposes a smarter way to run the show. Instead of solving the math puzzle from scratch every time, the system learns from a "Master Planner" (an expert computer program) and then uses that knowledge to make quick, smart guesses in real-time.

Here is how it works, broken down into simple parts:

1. The "Master Planner" and the Three Dictionaries

First, the researchers used a powerful, slow computer program (called a Genetic Algorithm) to solve hundreds of different drone missions perfectly. They didn't just save the flight paths; they broke those perfect flights down into three "dictionaries" or rulebooks:

The Mission Dictionary: This is like a master list of how to split up the work. If there are 50 houses to visit, this dictionary says, "Okay, Drone A takes the north side, Drone B takes the south side." It learns the best way to divide the job.
The Route Dictionary: Once the work is divided, this dictionary decides the order. "Drone A, visit House 1, then House 3, then House 2." It learns the most efficient loops.
The Motion Dictionary: This is the fine print. It describes how to fly between houses. Is it a straight shot? Do you need to slow down because of a tree? Do you need to swerve to avoid another drone? It turns complex physics into simple "motion words."

2. The "Active Inference" Brain

Once these three dictionaries are learned, the drones don't need the slow Master Planner anymore. They use a brain called Active Inference.

Think of this like a detective trying to solve a mystery. The detective has a "theory" of how the world works (based on the Master Planner's dictionaries). When the drones see something new—like a sudden new target or a moving obstacle—they ask: "Does this new situation fit our theory?"

The "Abnormality" Meter: The system calculates how "weird" or "abnormal" a potential action is compared to what the Master Planner would have done.
The Decision: If a drone considers a path that looks very different from the expert's rules, the "abnormality" meter goes high, and the drone rejects that path. If a path looks very similar to the expert's rules, the meter stays low, and the drone takes it.

This allows the swarm to make decisions instantly. They aren't calculating complex math; they are just checking, "Does this look like what a pro would do?"

3. Adapting to Changes (The "Replanning" Trick)

What happens if a new target appears while the drones are flying?

Old Way: Stop everything, recalculate the entire puzzle for all drones, and start over. (Too slow).
This Paper's Way: The system looks at the new target and asks, "Which drone should pick this up to keep the 'abnormality' low?" It then asks, "Where in that drone's current list should this new stop go?" Finally, it adjusts the flight path slightly. It's like a conductor adjusting the orchestra's sheet music on the fly without stopping the music.

4. The "Safety Net" (Filters)

Even with the smart dictionaries, real life is messy. Sensors get noisy, and wind blows. To handle this, the system uses Bayesian Filters (like EKF and PF).

The Analogy: Imagine you are driving a car in the fog. You know the road layout (the dictionary), but you can't see perfectly. The filter is like a co-pilot who says, "You think you're at the corner, but based on your speed and the fog, you're actually a few feet to the left."
The system uses this "co-pilot" to correct the drone's actual position before it makes a decision, ensuring they don't crash even if their sensors are a little fuzzy.

5. The Results

The researchers tested this in two ways:

Simulations: They ran thousands of computer missions. The new system flew just as efficiently as the slow "Master Planner" but was much faster at reacting to changes. It also flew much smoother and more steadily than other learning methods (like Q-learning), which tended to be jittery and erratic.
Real Flights: They tested it with real drones flown by humans indoors. Even though human pilots make mistakes and fly in jerky ways, the system was able to "clean up" the data. It could look at a messy human flight, figure out the intended path, and correct its own predictions to stay on track.

Summary

In short, this paper teaches a swarm of drones to learn from an expert rather than trying to be a genius mathematician every second. By breaking the problem down into three levels (Who does what? In what order? How do they fly?) and using a "weirdness detector" to stay on track, the drones can adapt to new situations instantly, fly smoothly, and avoid crashes, all without needing a supercomputer to solve the whole puzzle every time a new target appears.

1. Problem Statement

The paper addresses the challenge of adaptive trajectory planning for Unmanned Aerial Vehicle (UAV) swarms in dynamic and uncertain environments. The core problem involves solving a Multi-Traveling Salesman Problem (MTSP) where a swarm must:

Allocate tasks: Distribute a set of target locations among multiple UAVs.
Sequence routes: Determine the optimal visiting order for each UAV.
Generate trajectories: Create collision-free, dynamically feasible paths.
Adapt online: React to environmental changes (e.g., new targets, obstacles) in real-time without re-running computationally expensive global optimization.

Limitations of Existing Approaches:

Deterministic/Metaheuristic Optimization (e.g., Genetic Algorithms): Provide high-quality offline solutions but lack real-time adaptability; any environmental change requires a full re-optimization.
Deep Reinforcement Learning (DRL): Offers adaptability but suffers from high sample complexity, poor generalization to out-of-distribution scenarios, and lack of interpretability.

2. Methodology

The authors propose an expert-guided, active-inference-inspired framework that transforms the combinatorial optimization problem into a hierarchical probabilistic inference problem. The methodology consists of two phases:

A. Offline Phase: Expert Demonstration & World Model Learning

Expert Generation: A Genetic Algorithm with Repulsive-force collision avoidance (GA-RF) generates high-quality expert demonstrations for various mission instances.
Hierarchical Symbolic Abstraction: The continuous expert trajectories are decomposed into three symbolic dictionaries:
- Mission Dictionary: Encodes how the global set of targets is partitioned among UAVs (Task Allocation).
- Route Dictionary: Encodes the ordered sequence of visits for each UAV (Route Sequencing).
- Motion Dictionary: Encodes local motion primitives (e.g., target attraction, obstacle avoidance) derived from physical potential fields (attractive/repulsive forces).
Probabilistic World Model: The system learns a hierarchical probabilistic model from these symbols. It estimates:
- Reference distributions for mission patterns.
- Transition probabilities from Mission $\to$ Route $\to$ Motion.
- This model captures the conditional dependencies of expert behavior (e.g., how a specific mission allocation induces a specific route order).

B. Online Phase: Active Inference & Adaptive Control

Instead of solving the MTSP from scratch, the UAV swarm performs online inference to select actions that minimize "abnormality."

Abnormality Minimization: For a given observation, the swarm evaluates candidate actions at each hierarchical level. It calculates the Kullback-Leibler (KL) divergence between the inferred posterior belief of the action and the expert-derived reference distribution.
- Low KL Divergence: The action is consistent with expert behavior.
- High KL Divergence: The action is "abnormal" or inconsistent.
Hierarchical Decision Making: The swarm selects the action sequence (Mission, Route, Motion) that minimizes the total weighted abnormality.
- Adaptation: When a new target appears, the system does not re-optimize globally. Instead, it evaluates which UAV should take the new target (Mission level) and where to insert it in the route (Route level) to minimize the resulting abnormality.
Filter-Assisted Control: To handle noise and uncertainty in continuous state estimation, Bayesian filters (EKF and Particle Filter) are integrated at the motion level. These filters predict continuous states and update the likelihood of motion symbols, ensuring smooth trajectory correction and robust collision avoidance.

3. Key Contributions

Hierarchical Symbolic World Model: The construction of a three-level (Mission, Route, Motion) symbolic model learned from expert demonstrations, which abstracts complex trajectories into reusable probabilistic patterns.
KL-Based Active Inference Mechanism: A decision-making framework that replaces full free-energy minimization with KL-divergence-based abnormality indicators. This steers the swarm toward expert-consistent behaviors while avoiding the computational cost of global optimization.
Unified Probabilistic-Symbolic Architecture: A system that merges expert knowledge, probabilistic inference, and hierarchical symbolic control, offering both interpretability and adaptability.
Filter-Assisted Motion Correction: The integration of EKF/Particle Filters with symbolic inference to handle noisy observations and ensure safe, continuous trajectory execution.

4. Results and Validation

The framework was evaluated through extensive simulations and real-flight experiments:

Simulation Results:
- Performance: The proposed method achieved mission completion times and total traveled distances comparable to the offline GA-RF expert planner.
- Stability: Compared to a modified Q-learning baseline, the proposed framework produced smoother, more stable trajectories with less oscillatory behavior.
- Adaptability: The system successfully adapted to unseen target layouts and dynamic changes (e.g., new targets appearing) without retraining or re-running the global optimizer.
- Structure Preservation: The learned world model preserved the structural regularities of expert planning (task allocation and route ordering) better than reinforcement learning baselines.
Real-Flight Validation (Indoor):
- Using data from human-piloted DJI Air 2S UAVs, the system demonstrated robustness against sensing noise and non-smooth human pilot behaviors.
- Belief Correction: The Bayesian update mechanism successfully corrected symbolic prediction errors caused by noisy observations, aligning the predicted trajectory with the observed reality.
- Mission Success: Despite local motion prediction errors due to noise, the mission completion success rate remained at 100%, proving the robustness of the high-level hierarchical inference.

5. Significance

This paper presents a significant shift in UAV swarm autonomy from repetitive optimization to structured inference.

Computational Efficiency: By restricting the search space to expert-consistent symbolic candidates, the method enables real-time replanning that is computationally intractable for traditional MTSP solvers.
Interpretability: Unlike "black-box" deep learning models, the hierarchical symbolic structure allows engineers to understand why a specific mission allocation or route was chosen (based on learned probabilistic transitions).
Robustness: The integration of active inference with Bayesian state estimation provides a robust mechanism for operating in uncertain, noisy, and dynamic environments, bridging the gap between theoretical planning and real-world execution.

In conclusion, the framework offers a scalable, interpretable, and adaptive solution for complex multi-UAV missions, effectively leveraging expert knowledge to guide autonomous decision-making in real-time.