crazyplan

Research demo highlighting the differentiable crazyflow simulator for Crazyflie drones. It optimizes an open-loop policy to fly a racing track, then exports and visualizes the resulting trajectory.

Setup

Requirements: Python 3.11+, uv (recommended), and the crazyflow dependency (included in this repo).

Checkoout the repo and initialize the submodules

 git submodule update --init --recursive

Install uv and the projct in editable mode:

uv sync
uv pip install -e .

CLI

All entrypoints live under crazyplan/cli/. Run them in module mode so imports resolve cleanly.

• Optimize a trajectory (reads configs/train/default.toml by default):

  uv run python -m crazyplan.cli.train --train-config configs/train/default.toml --plot --render

  Outputs checkpoints to the path in the train config (train.policy_output / train.checkpoint_dir), optional plots/renders.

• Export a trajectory from a trained checkpoint (uses train.policy_output by default):

  uv run python -m crazyplan.cli.export --train-config configs/train/default.toml --visualize none
  # or pick a specific checkpoint
  uv run python -m crazyplan.cli.export --train-config configs/train/default.toml --checkpoint outputs/checkpoints_window/trained_policy_step_000100.eqx

  Writes YAML/CSV under outputs/export (default), with optional plot/render (--visualize plot|render|both).

• Plot track geometry:

  uv run python -m crazyplan.cli.plot_track --track-config configs/tracks/drone_racing.toml

Configs

Train configs live in configs/train/*.toml. Track configs live in configs/tracks/*.toml.

Gallery