💻 computer science

Online Structure Learning and Planning for Autonomous Robot Navigation using Active Inference

This paper introduces AIMAPP, a self-supervised, ROS-compatible framework that unifies online mapping, localisation, and planning for autonomous robots in unfamiliar environments by leveraging Active Inference to dynamically build sparse topological maps and balance exploration with goal-directed navigation without relying on predefined maps or pre-training.

Original authors: Daria de Tinguy, Tim Verbelen, Emilio Gamba, Bart Dhoedt

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Daria de Tinguy, Tim Verbelen, Emilio Gamba, Bart Dhoedt

Original paper licensed under CC BY 4.0 (https://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 a robot dropped into a brand-new, unfamiliar building. It has no blueprints, no GPS, and no one to tell it where to go. Most robots would get confused, bump into walls, or get stuck in loops.

This paper introduces a new way for robots to navigate called AIMAPP. Think of AIMAPP not as a robot with a hard-coded map, but as a robot with a curious, adaptive brain inspired by how animals (like rats or humans) navigate the world.

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

1. The "Mental Sketch" (Topological Map)

Instead of trying to build a perfect, pixel-by-pixel 3D model of the entire building (which is like trying to draw every brick in a wall), AIMAPP builds a mental sketch.

The Analogy: Imagine you are walking through a new city. You don't memorize every square inch of the pavement. Instead, you remember key landmarks: "The coffee shop," "The park," "The library." You connect them with lines in your mind: "If I walk straight from the coffee shop, I hit the park."
How the Robot Does It: The robot creates a "topological map." It treats distinct places as "nodes" (dots) and the paths between them as "links" (lines). As it explores, it adds new dots to its sketch. It doesn't need to know the exact metric distance down to the millimeter; it just needs to know, "This place is connected to that place."

2. The "Curious Explorer" vs. The "Goal Seeker"

The robot has two competing drives, but it doesn't need a human to tell it which one to pick. It balances them automatically using a concept called Expected Free Energy.

The Analogy: Think of a dog in a park.
- Curiosity (Exploration): The dog smells a strange bush and wants to investigate to see what's there (reducing uncertainty).
- Goal (Exploitation): The dog sees its owner waving and wants to run straight to them (achieving a preference).
How the Robot Does It: AIMAPP calculates a single score for every possible move. If the robot is lost, the "curiosity" score is high, so it goes explore. If it knows where the goal is, the "goal" score is high, so it goes there. It switches between these modes seamlessly without needing a human to press a button.

3. Learning on the Fly (No Pre-Training)

Many modern robots need to be trained for months on thousands of hours of video data before they can work. AIMAPP is different.

The Analogy: Imagine a tourist who arrives in a foreign country. They don't have a guidebook. They learn the layout by walking, getting lost, and realizing, "Oh, that red building is the same one I saw earlier."
How the Robot Does It: The robot starts with zero knowledge. As it moves, it compares what it thinks it will see with what it actually sees.
- If it sees something new, it adds a new "dot" to its map.
- If it sees something familiar, it updates its belief about where it is.
- It learns the rules of the environment (like "turning left here leads to a dead end") just by trying and failing, then correcting itself.

4. Handling "Drift" (The Wobbly Odometer)

Robots often use wheel sensors to guess how far they've moved. But on carpet or sand, wheels slip, and the robot thinks it moved 1 meter when it actually moved 2. This is called "drift."

The Analogy: Imagine walking with your eyes closed. You count your steps to guess where you are. If you slip on a rug, your step count is wrong. But if you open your eyes and see a familiar landmark (like a specific painting), you instantly know, "Ah, I'm actually here, not where my step count said I was."
How the Robot Does It: AIMAPP relies on visual recognition more than wheel sensors. If the robot's wheels say it's in a hallway, but its camera sees a kitchen, it trusts the camera. It corrects its internal map based on what it sees, not just what it feels. This makes it very robust in messy environments like sandy parking lots or carpeted rooms where other robots fail.

5. The Results: How Did It Do?

The researchers tested AIMAPP in both computer simulations and real-world settings (warehouses, houses, and even a sandy parking lot).

Efficiency: It explored areas almost as efficiently as a human walking through the same space with a map in hand.
Robustness: In a large warehouse, it handled moving obstacles and slippery floors better than other advanced robots that rely on precise maps.
Failure: It did get stuck occasionally if the robot's internal map became too "messy" (e.g., if the robot drifted so far it thought it was in a new place when it was actually in an old one), but it recovered much better than standard robots.

Summary

AIMAPP is a navigation system that treats a robot like a curious animal. Instead of following a rigid script or a pre-loaded map, it builds a mental sketch of the world as it goes, balances its desire to explore with its desire to reach a goal, and constantly corrects its position by recognizing familiar sights. It works without prior training, making it ready to be dropped into any new environment and start figuring things out immediately.

Technical Summary: Online Structure Learning and Planning for Autonomous Robot Navigation using Active Inference

Problem Statement

Autonomous mobile robots operating in unfamiliar environments face the challenge of simultaneously localizing, mapping, and planning without reliance on pre-defined maps, extensive offline training, or human supervision. Existing approaches often suffer from specific trade-offs:

Metric SLAM systems (e.g., ORB-SLAM3, FAST-LIO2) provide accurate pose estimation but are prone to drift in large or unstructured environments and lack intrinsic exploration policies.
Path planners often incur high computational costs or rely on greedy heuristics (e.g., frontier-based exploration) that produce redundant backtracking.
Learning-based approaches (e.g., Neural-SLAM, ViKiNG, NoMaD) improve flexibility but typically require extensive offline data collection, pre-training, or teleoperated demonstrations in the target environment, limiting adaptability to genuinely novel settings.

The core challenge addressed is creating a unified system that performs online structure learning and planning in a fully self-supervised manner, resilient to sensor noise, odometric drift, and environmental changes, without requiring pre-training or hand-engineered priors.

Methodology: AIMAPP Framework

The authors propose AIMAPP (Active Inference MAPping and Planning), a framework that unifies mapping, localization, and decision-making within a single generative model based on the Active Inference Framework (AIF).

1. State Representation and Cognitive Mapping

AIMAPP maintains a sparse topological map where the latent state is decomposed into two components:

$s_t$ (Discrete State): A topological index representing a distinct node in the graph (a "place").
$p_t$ (Continuous Metric Position): Local 2D metric coordinates used for spatial reasoning.
This decomposition mirrors biological navigation systems (separating place-cell identity from grid-cell path integration). The map grows adaptively; new nodes are created when motion predictions exceed the "influence radius" of existing states, determined by free-energy comparisons. Observations are stored as 360° panoramic images (or abstract sensor data) associated with nodes.

2. Generative Model and Inference

The system operates as a Partially Observable Markov Decision Process (POMDP). The generative model includes:

State transitions ( $B_s$ ): Likelihood of moving between topological states.
Position transitions ( $B_p$ ): Likelihood of moving between metric positions.
Observation likelihoods ( $A_o$ ): Probability of observing a specific panorama at a given state.
Position likelihoods ( $A_p$ ): Probability of occupying a metric position at a given state.

Localization is belief-driven. The agent infers its current state by minimizing Variational Free Energy (VFE), combining predicted motion with sensory evidence. This allows the system to remain robust to odometric drift; if drift occurs, the agent prioritizes consistency between predicted and observed outcomes, re-localizing by seeking familiar observations rather than relying solely on raw odometry.

3. Planning and Decision-Making

Action selection is driven by the minimization of Expected Free Energy (EFE) ( $G$ ). Unlike traditional reinforcement learning where exploration and exploitation are separate objectives, AIMAPP derives both from the same EFE term:

Epistemic Value (Exploration): Expected information gain to reduce uncertainty about the environment (states and parameters).
Pragmatic Value (Exploitation): Expected alignment with preferred observations (e.g., reaching a target or avoiding collisions).

Candidate policies are evaluated using Monte Carlo Tree Search (MCTS) to handle the combinatorial complexity of action spaces. The EFE for a policy $\pi$ is decomposed into terms representing information gain on states/parameters, expected collision avoidance, and utility terms for observations, positions, and states.

4. Online Learning and Map Expansion

The system updates its internal parameters continuously. Map expansion is triggered when an "expanded model" (incorporating a new state) yields a lower free energy ( $\Delta F < 0$ ) compared to the current model. Transition models ( $B_s$ ) and observation likelihoods ( $A_o$ ) are updated using Dirichlet pseudo-counts based on experienced vs. predicted transitions, allowing the agent to refine its world model from sensorimotor feedback without pre-training.

Key Contributions

Unified Active Inference Framework: AIMAPP integrates mapping, localization, and planning into a single generative model, eliminating the need for separate modules or hand-arbitrated trade-offs between exploration and goal-reaching.
Fully Self-Supervised Online Learning: The agent builds and refines its topological map from the first interaction without pre-training, offline data collection, or teleoperated demonstrations.
Robustness to Drift and Noise: By prioritizing belief consistency over raw odometry, the system tolerates sensor drift and operates in environments with perceptual aliasing and moderate changes.
Modularity and Agnosticism: The architecture is sensor-agnostic (supporting RGB, depth, LiDAR) and platform-agnostic (ROS-compatible), separating the cognitive layer from perception and motion control.
Scalable Evaluation: The system is evaluated in environments up to 325 m² (real and simulated), a scale previously unaddressed by Active Inference navigation systems.

Experimental Results

The authors evaluated AIMAPP against state-of-the-art baselines (Frontier-based, GBPlanner, FAEL) and human teleoperation in four simulated and three real-world environments.

Exploration Efficiency: In simulated environments, AIMAPP achieved coverage efficiency comparable to or exceeding state-of-the-art methods (FAEL, GBPlanner). Specifically, it reached 90.3% of the coverage efficiency of human teleoperation on average. In real-world trials, AIMAPP consistently outperformed Frontier-based exploration in terms of coverage speed and success rate.
Robustness: AIMAPP successfully operated in conditions where odometry-based baselines (e.g., Frontiers with Nav2) failed, such as sandy parking lots and carpeted floors with significant drift. It demonstrated resilience to moved obstacles in simulation and sensor noise in real-world settings.
Goal-Seeking: The system achieved goal-reaching performance within 74% of human-teleoperated performance across all tested environments without task-specific re-tuning.
Success Rates: AIMAPP achieved a 79% success rate in obtaining five successful exploration trials per environment, outperforming Frontiers (52%) and FAEL (48%), though slightly below GBPlanner (87%). Failures were primarily attributed to unrecoverable map overlaps due to severe drift in large unexplored areas or physical collisions with undetected obstacles.

Significance and Claims

The paper claims that AIMAPP demonstrates that Active Inference, originally a theory from computational neuroscience, can be successfully adapted into a modular, robust navigation strategy for real robots. The significance lies in treating navigation as hypothesis testing: predictions about future states are generated, evaluated through action, and revised based on "surprise" (prediction error).

The authors emphasize that AIMAPP offers a modular, self-supervised solution for scalable navigation in unstructured settings. By unifying exploration and goal-reaching within a single objective (minimizing Expected Free Energy), the system avoids the need for task-specific tuning or pre-existing knowledge of the environment. The work represents a step toward autonomous agents that can discover the structure of their environment from the first interaction, operating effectively in dynamic, unstructured, and unfamiliar spaces where traditional metric mapping or pre-trained learning models struggle.