Spatial and Temporal Hierarchy for Autonomous Navigation using Active Inference in Minigrid Environment

Autonomous navigation has evolved from classical SLAM algorithms (e.g., GMapping, Cartographer) to deep learning-based approaches (e.g., DQN, A3C). While these methods excel in structured environments, they struggle with scalability, dynamic changes, and environment generalization. Cognitive science insights reveal that humans use hierarchical spatial representations and path integration, enabling flexible and robust navigation. Recent research incorporates neural networks with probabilistic models, but often lacks the ability to reason over multiple scales or handle environment ambiguity effectively. Integrating these biological principles with probabilistic inference frameworks offers a promising direction for advancing autonomous navigation.

Current navigation systems face challenges in complex, dynamic environments, including high computational costs, poor scalability, and difficulty in environment understanding. Traditional SLAM approaches are limited by environmental complexity and sensor noise, while deep learning models often require extensive training data and lack interpretability. The core problem is developing a scalable, robust model that can learn environment structure, perform long-term planning, and adapt to environmental changes without relying heavily on labeled data or computationally intensive processes.

This work introduces a hierarchical active inference framework that models environment understanding at multiple abstraction levels. Key innovations include: 1) a three-layer generative model mimicking human spatial reasoning; 2) integration of pixel-based visual perception with Bayesian path planning; 3) use of variational inference for efficient environment mapping; 4) a novel combination of cognitive maps, allocentric, and egocentric models operating at different timescales. These innovations enable autonomous agents to learn environment structure, plan efficiently, and adapt dynamically, addressing limitations of existing SLAM and deep learning methods.

• �� Construct a three-layer hierarchical generative model: cognitive map (top layer, coarse spatial structure), allocentric model (local environment details), egocentric model (motion and sensory prediction).
• �� Use variational inference to estimate latent states, updating beliefs based on visual observations and actions.
• �� Employ Bayesian path inference by calculating expected free energy (EFE), balancing information gain and goal utility.
• �� Integrate pixel-based visual inputs for environment representation, enabling learning without explicit labels.
• �� Simulate environments in MiniGrid, evaluating goal success rate, path efficiency, and robustness.
• �� Conduct ablation studies to assess each layer's contribution and the model's adaptability to environment changes.

The model was tested in MiniGrid environments with varying maze complexities, measuring goal success rate, path length, exploration efficiency, and robustness against environmental noise. Baseline comparisons included DQN and single-layer active inference models. Hyperparameters such as learning rate, inference horizon, and planning depth were tuned for optimal performance. Results showed a 20% improvement in goal success, 30% reduction in path length, and increased resilience to environment perturbations. The experiments validated the model's ability to build internal environment maps, handle aliasing, and perform long-term planning.

The hierarchical model achieved 85% goal success in complex mazes, outperforming RL baselines by over 20%. Path planning was 30% faster, with exploration efficiency increased by 25%. The model demonstrated strong environment understanding, constructing internal cognitive maps that resist aliasing and adapt to dynamic changes, significantly surpassing traditional SLAM and flat models in robustness and scalability.

Potential applications include autonomous robots in warehouses, search-and-rescue drones, and self-driving vehicles operating in unfamiliar or dynamic environments. The model's ability to learn environment structure from visual input and plan over multiple scales makes it suitable for real-world deployment, especially where environment labels are unavailable or costly to obtain. It can also be integrated with multi-sensor systems for enhanced perception.

The model's real-time performance is limited in highly dynamic environments due to computational overhead in updating hierarchical maps. Its reliance on environment assumptions may lead to errors in highly unpredictable scenarios. Additionally, high computational costs pose challenges for deployment on resource-constrained platforms. Future work should focus on optimizing inference efficiency, extending to multimodal perception, and validating in real-world robotic systems.