FlyDreamer: Learning to Fly FPV in Dreams

Note: This project is forked and modified from the official DreamerV3 repository by Danijar Hafner. Full credit for the base code structure and DreamerV3 algorithm goes to the original authors and Google DeepMind.

---

Overview

FlyDreamer trains a reinforcement learning agent to fly an FPV (First-Person View) drone in simulation using world model-based learning. The agent learns entirely from imagination—dreaming trajectories in a learned world model before executing in the real environment.

Supported Simulator: Custom Colosseum Fork (based on Colosseum, which is itself a fork of Microsoft AirSim)

Quick Start

Manual Controller Test

Test the FPV environment with a physical controller (TX16S and other joysticks supported):

python test_manual_control.py

Training

Train an FPV navigation agent:

python dreamerv3/main.py \
  --configs fpv \
  --logdir logs/fpv_run

World Model Pretraining (Experimental)

Pretrain the world model on demonstration data before policy learning:

python dreamerv3/main.py \
  --configs fpv pretrain_wm \
  --logdir logs/pretrain \
  --demo_data_dir demo_data/chunks

TSSM Transformer Dynamics (Experimental)

Use a Transformer-based state-space model instead of the default RSSM:

python dreamerv3/main.py \
  --configs fpv \
  --agent.dyn.typ tssm \
  --logdir logs/tssm_run

Development Status

This project is under active development and experimentation. Documentation, features, and APIs may change frequently. More detailed documentation will be added in the future.

References

Core Algorithm

• DreamerV3: Hafner, D., Pasukonis, J., Ba, J., & Lillicrap, T. (2025). Mastering diverse control tasks through world models. Nature. [Paper] [Website] [Code]

• DreamerV2: Hafner, D., Lillicrap, T., Norouzi, M., & Ba, J. (2021). Mastering Atari with Discrete World Models. ICLR. [Paper]

Related Drone Learning Work

• SkyDreamer: Wu, X., et al. (2024). SkyDreamer: Autonomous Drone Racing with World Models. [Paper]

• Dream to Fly: Zhou, P., et al. (2024). Learning to Fly FPV Drones in Virtual Reality. [Paper]

• Learning to Fly in Seconds: Kaufmann, E., et al. (2023). Champion-level drone racing using deep reinforcement learning. Nature. [Paper]

Simulators

• Colosseum (Custom Fork): Modified Colosseum with FPV-specific enhancements. [GitHub]

• Colosseum: CodexLabs fork of AirSim with continued maintenance. [GitHub]

• AirSim: Microsoft Research drone and car simulation platform. [GitHub] [Docs]

Citation

If you use this code, please cite the original DreamerV3 paper:

@article{hafner2025dreamerv3,
  title={Mastering diverse control tasks through world models},
  author={Hafner, Danijar and Pasukonis, Jurgis and Ba, Jimmy and Lillicrap, Timothy},
  journal={Nature},
  year={2025},
  publisher={Nature Publishing Group}
}

License

This project follows the license of the original DreamerV3 repository.