Drone navigation excels with action-based monocular vision model
Skytopia: Monocular Drone Navigation with Action-Conditioned Latent World Models
RoboticsArtificial Intelligence
Summary
Flying a drone using just one front-facing camera is hard because it’s tough to judge distances and scale. The authors created a new system called Skytopia that helps drones understand how what they see changes when they move. Instead of predicting exact images all the time, Skytopia focuses on simplifying the representation of the environment based on the drone’s actions. This approach lets drones navigate better in different settings, like indoors or outside, without needing extra training. Their tests show this method is more successful and efficient than previous ones.
What this means in practice
- •For drone software engineers: Develop navigation software that reliably guides drones using only a single forward camera across diverse environments without retraining.
- •For robotics platform developers: Integrate efficient latent world models to reduce computation during autonomous navigation, improving real-time control in robotic platforms.
Authors
Yuhang Zhang, Rangya Zhang, Yujing Shang, Zhuoyuan Yu, Weiying Wang, Steven Yang, Qingsong Yan, Chao Yan, Mir Feroskhan
Abstract
Monocular drone navigation requires reaching a goal in an unseen environment from a single forward-facing camera, which offers few cues for depth and scale. World models address this by modelling how observations evolve under actions, but they are built to be executed: the prediction is produced at deployment and fed back into action generation at every control step. We argue that what a policy needs from a world model is not the prediction but the representation required to produce it: in flight the executed action explains almost all of the change between observations, so prediction reduces to reprojecting a static scene under a known displacement. We therefore introduce skytopia, a policy built on an action-conditioned latent world model, and the 3D Gaussian Splatting platform on which it is trained. A forward objective predicts the representation of the next observation from the intended motion, and an inverse objective recovers that motion from the predicted transition. Because the prediction never reaches action generation, the predictor is discarded and one policy serves point-goal, image-goal, and goal-free navigation. Simulation experiments show that skytopia outperforms every baseline under all three specifications, attaining 57.8%, 66.0%, and 49.0% success rate, while discarding the predictor removes 59.4% of the inference cost. The same policy is subsequently deployed on a physical drone without fine-tuning and reaches goals in indoor, open outdoor, and woodland environments.