Zero-shot Structure Learning and Planning for Autonomous Robot Navigation using Active Inference

TL;DR

A ROS-compatible navigation system that is sensor and robot-agnostic, capable of integrating with diverse hardware configurations, and supports both exploration and goal-directed navigation without any pre-training is implemented.

Abstract

Autonomous navigation in unfamiliar environments requires robots to simultaneously explore, localise, and plan under uncertainty, without relying on predefined maps or extensive training. We present a biologically inspired, Active Inference-based framework, Active Inference MAPping and Planning (AIMAPP). This model unifies mapping, localisation, and decision-making within a single generative model. Inspired by hippocampal navigation, it uses topological reasoning, place-cell encoding, and episodic memory to guide behaviour. The agent builds and updates a sparse topological map online, learns state transitions dynamically, and plans actions by minimising Expected Free Energy. This allows it to balance goal-directed and exploratory behaviours. We implemented a ROS-compatible navigation system that is sensor and robot-agnostic, capable of integrating with diverse hardware configurations. It operates in a fully self-supervised manner, is resilient to drift, and supports both exploration and goal-directed navigation without any pre-training. We demonstrate robust performance in large-scale real and simulated environments against state-of-the-art planning models, highlighting the system's adaptability to ambiguous observations, environmental changes, and sensor noise. The model offers a biologically inspired, modular solution to scalable, self-supervised navigation in unstructured settings. AIMAPP is available at https://github.com/decide-ugent/AIMAPP.

Figures (19)

• Figure 1: Our model coarsely localises itself based on local motion and accumulated observations, remembering in a topological map where it has been and could potentially go. Planning follows Active Inference Principles, considering where the agent's preferences lie (whether to gather information about the surroundings or reach a specific objective). Visited states (up to time $t$ presented with coloured circles on the topological map) have a visual observation and a known position, while unvisited states (blank circles with dotted edges) hold a probable position and no observation. We plan through consecutive nodes in the topological graph to determine the next motion.
• Figure 2: POMDP of our model. $Cat$ stands for Categorical. The model integrates states ($s_t$), positions ($p_t$), and observations ($o_t$) over time, guided by policies ($\pi$) and expected free energy ($G$). The categorical distributions define transition and observation likelihoods: $A_p$ (position likelihoods), $A_o$ (observation likelihoods), $B_p$ (position transitions), and $B_s$ (state transitions). This structure underpins the inference scheme in Equation \ref{['eq2']}, enabling the agent to infer hidden states and positions from sensory observations and prior beliefs.
• Figure 3: Influence of a node at position (0,0) on adjacent node creation for an influence radius of 0.5 m and 12 discrete headings spanning $360^\circ$. Red circles represent the influence radius of each newly created node. The blue dashed circle marks the robot’s radius, including a padding term to account for its physical size (important near walls). The 12 black dots correspond to the midpoints of the 12 possible action directions. Dots with a black aura indicate valid positions where a new node can be created while respecting the minimum distance constraint (red radius) of adjacent nodes. Dots without an aura represent invalid positions (too close to an existing node) and could instead be created farther away (e.g., at 1 m). This arrangement allows the agent to maintain open junctions for future node expansion.
• Figure 4: schematic of what could be the topological map in a simple 2-room environment. Dots are states, or nodes, each state containing a $360^\circ$ panorama obtained through stitching images together and associated with a position. Links between nodes are plausible transitions.
• Figure 5: Overview of the system architecture. In grey, we have modules that any ROS-compatible solution can replace. Modules interact through belief propagation. Inferring and planning (localisation, mapping and action selection) rely on the AIF framework. The perceptual and motion planning still use traditional approaches. Believed odometry takes precedence over sensor odometry. Preferences (goal) are expected from the user if we want to reach a target observation.