Ethgar/skyjepa-quadrotor-sim

Part of Ethgar's robot lab notebook: the training data of the SkyJEPA 5k reference model. The code repository is private for now.

SkyJEPA Quadrotor Simulation Trajectories

Domain-randomized closed-loop quadrotor trajectories generated entirely in simulation for training the SkyJEPA latent-dynamics world model. This is the training corpus for a clean-room reimplementation of SkyJEPA (Rao et al., 2026, arXiv:2606.23444).

• 5,000 trajectories total
• Train: 4,000 (80 shards) · Val: 500 (10 shards) · Test: 500 (10 shards)
• No wind, no vision — pure state/action/domain-parameter trajectories.

How it was generated

A from-scratch RK4 rigid-body quadrotor simulator (SO(3) exp-map attitude, motor lag, drag) flies reference trajectories under 500-parameter domain randomization (mass, inertia, thrust/torque coefficients, drag, motor time constants, etc.). References come from Gaussian-process paths and differential-flatness maneuvers. ~25–30% of trajectories are tracked by an MPPI controller and the rest by a geometric SE(3) controller, giving a mix of smooth and aggressive closed-loop behavior. Diverged rollouts are rejected. Everything runs at 20 Hz. No real flight data is included.

Schema

Each split is a set of compressed .npz shards named {split}_{id}.npz. Every shard holds fixed-length trajectories:

array	shape	dtype	meaning
states	(n, 201, 18)	float32	18-dim state per step: position(3), velocity(3), rotation matrix(9), body angular rate(3)
actions	(n, 200, 4)	float32	4-dim action per step (normalized motor commands)
eta	(n, 13)	float32	per-trajectory domain-randomization parameters η

There are N states and N-1 actions per trajectory. Splits are disjoint at the shard level (deterministic 80/10/10 by shard id).

Usage

pip install "skyjepa[hf] @ git+https://github.com/edgarmoreaualix/LeDrone.git#subdirectory=skyjepa"
python scripts/download_dataset_hf.py --repo-id Ethgar/skyjepa-quadrotor-sim --out-dir data/full

from skyjepa.data.dataset import TrajectoryWindows
ds = TrajectoryWindows("data/full", split="train", history=10, horizon=20)
state_hist, action_hist, future_states, future_actions = ds[0]

Then train the full model on a GPU:

python scripts/run_train.py --profile full --stage both --data-dir data/full

License & credit

MIT. Method credit belongs to the original SkyJEPA authors (Rao et al., 2026, arXiv:2606.23444); this dataset is from an independent clean-room reimplementation of their simulation pipeline and is not affiliated with or endorsed by them.