Navigation and Exploration with Active Inference: from Biology to Industry

TL;DR

Addresses robotic navigation in unknown, dynamic environments without pre-training by unifying localisation, mapping and planning within a biologically inspired Active Inference framework. The system jointly infers position $p$ and latent state $s$, updates a growing generative model under free energy minimization, and expands the map when $DeltaF < 0$, guided by $A_p$, $B_s$, and $A_o$ in the Expected Free Energy objective. Implemented in ROS2 as a modular architecture with perception and control integration, it remains robust to sensor drift and kidnapping. Empirical results in 2D and 3D simulated environments and real indoor settings show competitive exploration efficiency against Frontiers and Gbplanner, validating data-efficient, adaptive navigation without pre-training.

Abstract

By building and updating internal cognitive maps, animals exhibit extraordinary navigation abilities in complex, dynamic environments. Inspired by these biological mechanisms, we present a real time robotic navigation system grounded in the Active Inference Framework (AIF). Our model incrementally constructs a topological map, infers the agent's location, and plans actions by minimising expected uncertainty and fulfilling perceptual goals without any prior training. Integrated into the ROS2 ecosystem, we validate its adaptability and efficiency across both 2D and 3D environments (simulated and real world), demonstrating competitive performance with traditional and state of the art exploration approaches while offering a biologically inspired navigation approach.

Figures (8)

• Figure 1: Left: Factor graph of the POMDP generative model, showing transitions from past to future (up to time $t+1$). Known observations (blue) inform current latent states. Future actions follow policy $\pi$, influencing inferred positions and states (orange), and generating predictions of future observations (grey). The agent's position $p_t$ is determined by $p_{t-1}$ and the selected policy, while the latent state $s_t$ is inferred from $o_t$, $p_t$, and $s_{t-1}$. Transitions are parametrised by $B$ matrices, and $A$ matrices encode the likelihood of observations given latent states. Right (a–c): Three ways of structuring the world, progressing from the most common (a) with a given static world dimension, to (b) a growing state learning given a new observation, to (c) the structure learned by our proposed model given expected motions.
• Figure 2: Our results (second column) compared to L.-E. Martinet & al's (third column) tolman_maze2_redone. In our study, the agent's flow paths towards the objective (top of the map) are shown, with re-planning occurring when the desired path is blocked. The varying colour gradient of the lines indicates the frequency of selection for each path over all agents. A sequence is read horizontally, a) is the maze without obstacles, and b) and c) illustrate obstacles at points A and B, respectively. The occupancy grid maps demonstrate the learning of maze topology by simulated agents, initially without obstacles, showing a significant preference for Route 1. When a block is introduced at point A, the animals predominantly choose Route 2. With an obstacle placed at point B, the animals mainly opt for Route 3.
• Figure 3: Overview of the system architecture. Modules interact through belief propagation, Inferring and planning (localisation, mapping and action selection) rely on the Active Inference framework. The perceptual and motion planning still use traditional approaches. Believed odometry takes precedence over sensor odometry. Preferences are expected from the user if we want to reach a target observation. Red contours highlight newly added or modified modules for Real-world navigation.
• Figure 4: Final map of exploration in a) Amazon simulated warehouse, b) a real-world environment. Coloured points signify visited locations, where the same colour attributions mean the same observation. The thickness of the lines depicts the agent's believed probability of transitioning between two states given an action.
• Figure 5: Lidar Coverage of a 280m$^2$ warehouse over the distance travelled (m) with our model, Frontiers and Gbplanner over five runs each.