Abstract
Deep Active Localization (DAL) is an end-to-end differentiable framework that enables robots to determine their pose within a map by executing intentional, informative actions. By combining a hierarchical deep learning perception module with a reinforcement learning-based planning module, the system resolves positional ambiguity efficiently. Trained entirely in simulation with aggressive domain randomization, DAL demonstrates robust zero-shot transfer to real-world hardware, outperforming traditional hand-tuned localization methods in both speed and accuracy.
Introduction
Localization—the ability of a robot to determine its coordinates on a map—is a cornerstone of autonomous navigation. While passive localization relies on incoming sensor data, active localization empowers the robot to choose actions (e.g., moving toward a unique landmark) to disambiguate its position rapidly. Traditional approaches like Adaptive Monte Carlo Localization (AMCL) often depend on hand-engineered models and heuristics that are brittle in noisy, real-world environments. Bridging the "reality gap" between simulation and physical deployment remains a significant challenge for learning-based methods. This paper introduces Deep Active Localization (DAL), a framework designed to overcome these limitations through a unified, differentiable architecture.
Methodology
The DAL framework consists of two primary components: a perception module and a planning module, integrated into a hierarchical structure.
Perception Module
The perception module utilizes a Convolutional Neural Network (CNN) to predict the likelihood of the robot's pose. To balance computational efficiency with precision, a multi-scale approach is employed:
- Level 0 (Coarse): Generates a low-resolution belief map used by the planning module for high-level decision-making.
- Level 1 (Fine): Focuses on the most probable regions identified by the coarse map to provide high-precision estimates for motor control.
Planning and Training
The planning module uses the Advantage Actor-Critic (A2C) algorithm to learn a policy that outputs high-level actions (Left, Right, Straight) based on the current belief and environment maps. A key innovation is the reward function, which uses the probability mass at the true pose, providing a denser training signal than traditional entropy-based metrics. To ensure real-world robustness, the authors utilized Domain Randomization (DR), intentionally introducing noise and distortions during training on 10,000 synthetic maps.
Results and Evaluation
The system was evaluated on physical robots (JAY and Turtlebot) in diverse laboratory settings.
Zero-Shot Transfer and Accuracy
DAL achieved successful zero-shot transfer, operating on real hardware immediately after simulation training. In real-world tests, the model consistently outperformed standard scan-matching baselines, achieving a higher "hit rate" and lower Wasserstein distance (belief error).
Efficiency and Robustness
The RL-based policy is highly efficient, computing the next action in approximately 1 millisecond, compared to over a second for traditional Active Markov Localization (AML). Furthermore, the learned likelihood model proved resilient to environmental changes, such as moved or added obstacles, where traditional methods often failed.
Conclusion
Deep Active Localization provides a robust, efficient, and transferable solution for autonomous robot positioning. By leveraging hierarchical perception and reinforcement learning, the system moves beyond the limitations of hand-tuned heuristics. Future work aims to extend this framework toward "map-less" localization, potentially utilizing abstract topological maps or hand-drawn sketches to navigate complex environments.
Summary updated