Reproduce “Goal-Guided Transformer-Enabled Reinforcement Learning for Efficient Autonomous Navigation” on OpenResearch

About this paper

Wenhui Huang, Yanxin Zhou, Xiangkun He, Chen Lv

Researchers from Nanyang Technological University developed Goal-guided Transformer-enabled Reinforcement Learning (GTRL), an approach that integrates physical goal states into a novel Vision Transformer (GoT) architecture for perception, enhancing data efficiency in deep reinforcement learning for mapless autonomous navigation. The method demonstrates superior navigation performance, robustness, and interpretability in both simulation and real-world environments, achieving 100% success rates with zero human engagements in real-world multi-goal tasks.

Key insights

• Directly coupling goal information with scene perception via a Transformer's attention mechanism effectively guides feature extraction towards task-relevant elements, significantly improving DRL data efficiency.
• The GoT's self-attention mechanism enhances interpretability by explicitly highlighting goal-oriented image patches, offering insights into the agent's decision-making process.
• Pre-training the goal-guided perception module with expert priors (Imitation Learning) provides a beneficial initialization that accelerates DRL convergence and overall learning stability.

Summary by alphaXiv.