Published in arXiv, 2026
NavGen composes visual generative models into a scalable data engine, generating a 400K-episode vision-language navigation dataset with long-tail coverage and 75% zero-shot real-world success.
Published in arXiv, 2026
A fully automated pipeline for collecting diverse and scalable UAV navigation data, using generative world models to create high-fidelity scenes, autonomous exploration to expand coverage, and dynamically feasible planning for trajectory generation.
Published in CoRL, 2026
A navigation system that uses BEV images as global priors and transfers the world-understanding ability of image generation models to long-range embodied navigation.
Published in RA-L, 2026
A zero-shot 3D navigation framework that leverages video generation models to predict future observations and guide the robot.
Published in RA-L, 2025
This work presented a method for autonomous high-speed flight in complex environments using only monocular RGB cameras, trained within 3DGS.
Published in RA-L, 2025
A reinforcement learning approach for reactive aerobatic maneuvers, enabling drones to perform agile tasks in constrained spaces.