Optimizing Dynamic Learning from Learned Hallucination (Dyna-LfLH) using a Transformer-based Motion Planner for Agile Navigation in Dynamic Environments

Autonomous robot navigation in dense, dynamic environments is a fundamental challenge. Classical planners struggle with unpredictable obstacles due to computation limits, while learning-based methods are hampered by needing high-quality demonstrations for Imitation Learning or exploration inefficiencies in Reinforcement Learning. Self-supervised frameworks like Dynamic Learning from Learned Hallucination (Dyna-LfLH) offer a promising solution by generating training data without costly demonstrations. However, current frameworks are limited by less expressive motion planners. The original Dyna-LfLH framework utilized a two-layer Recurrent Neural Network (RNN) that struggles to model complex, long-range spatio-temporal dependencies. This work replaces the RNN planner with an autoregressive, Transformer-based architecture. Our model uses a multi-layer causal Transformer encoder with GELU activations to process sequences of command embeddings fused with features from a three-layer convolutional LiDAR encoder. This multi-head self-attention architecture enables a crucial shift from myopic single-command prediction to multi-step trajectory forecasting, providing foresight for smoother navigation. We trained our model on an Imitation Learning dataset of 53,000 dynamic and open-space augmentation (OSA) examples generated via the Dyna-LfLH algorithm. The model learns to predict a 3-step command trajectory of linear and angular velocities (v,ω) from historical 2D LiDAR scans, prior commands, and a target position vector, minimizing Mean Squared Error (MSE). The final model was evaluated on a Jackal robot in the DynaBARN simulation testbed, achieving an improvement in collision avoidance compared to the original RNN-based implementation. This research presents a Transformer-based framework for motion planning in dynamic, unpredictable environments, improving the safety and reliability of autonomous navigation.