Quadrotors operating in cluttered spaces must avoid both stationary objects and moving threats, but most navigation policies are designed for only one of those settings. TransNav, a transformer-based end-to-end method, targets both with a single architecture that fuses perception and planning. It encodes point clouds in patches, combining local details with global context, and directly outputs a short action sequence, executing the first action.
The policy is trained without expert demonstrations. Instead, an unsupervised scheme uses a lightweight point-mass dynamics model to propagate numerical gradients and optimize trajectory-level costs over the predicted sequence. This simplifies training while keeping the policy compact enough for onboard inference.
In simulation and real-world flights, TransNav performed strongly in dense static environments and remained competitive in dynamic scenarios, with low computational overhead. The work appears in IEEE Robotics and Automation Letters.