SkyJEPA is a research project for learning latent dynamics models for aerial robots from principled simulation data and deploying them for zero-shot sim-to-real control.
The method combines JEPA-style latent prediction with a physics-inspired probing stage based on differentiable kinematic integration, enabling sampling-based control on real drone trajectories.
Code release: the training, evaluation, and deployment code will be released soon.
SkyJEPA follows a four-stage pipeline:
- Sim data synthesis: generate structured drone dynamics data in simulation.
- Latent dynamics model: learn a JEPA-style predictive representation from observations, states, and actions.
- Physics-inspired probe: decode future state structure through differentiable kinematic integration.
- Zero-shot sim-to-real: use the learned model for sampling-based control on real-world drone flights.
No-payload flight 01 |
Payload flight |
No-payload flight 02 |
The original high-resolution teaser videos are also available in assets/.
This repository currently hosts the project overview and visual assets. Source code, pretrained models, dataset instructions, and reproduction scripts will be released soon.
This project is released under the Apache License 2.0. See LICENSE for details.