Event

TransNav: A Transformer-Based End-to-End Obstacle Avoidance Navigation Method for Quadrotors

1 reports · 1 sources

Jianghan Zhu et al. proposed TransNav, a Transformer-based end-to-end navigation method for quadrotors, for obstacle avoidance in environments with mixed static and dynamic obstacles. The method integrates perception and navigation into a single architecture, employs patch-based point cloud encoding, and directly predicts a finite-horizon action sequence and executes the first action. For training, the authors introduced a lightweight point-mass dynamics model and directly optimized a trajectory-level cost over the predicted sequence through numerical gradient propagation, enabling unsupervised training without expert demonstrations. The paper states that simulation and real-world experiments show that the method performs strongly in dense static environments, has competitive obstacle avoidance capability in dynamic scenes, and maintains low onboard inference overhead.