From Perception to Action: Transformer-Enhanced Deep Reinforcement Learning for Autonomous Robot Navigation
Belabed Abdelkader, Nechadi Emira, Essounbouli Nadjib · 2025
Autonomous robot navigation is still a challenging task that requires strong capabilities in perception, decision-making, and control. Although multi-layer networks (MLP) and long short-term memory networks (LSTM) have shown promising results, they still face issues with long-term historical data, making the robot unable to cope with challenging and dynamic environments to reach the goals. In this paper, we present Transformer-Enhanced Deep Reinforcement Learning, a novel framework that leverages the multi-head attention mechanism of transformers to improve the navigation of autonomous robots. Our approach integrates transformers into a DRL pipeline, enabling the robot to capture long-range dependencies in sensory input effectively. We assess the performance of our model within a simulated environment using the GAZEBO simulator and ROS2 Humble framework, comparing it against GRU and MLP baselines. Experimental results demonstrate that the Transformer-based DRL model achieves superior performance, with higher success rates, smoother trajectories, and improved generalization to unknown environments.