Ethgar/skyjepa-mission-5k

SkyJEPA Quadrotor Simulation Trajectories — dataset v2 (mission-mix)

Domain-randomized closed-loop quadrotor trajectories generated entirely in simulation for training the SkyJEPA latent-dynamics world model. Version 2 trades the paper's wide airframe randomization for a narrow, bench-measurable airframe and adds wide external disturbances — the model spends its capacity learning to reject wind, payload changes, air-density shifts and battery sag rather than absurd airframes.

• 5,000 trajectories total
• 2,920 mission-shaped / 2,080 aggressive (58% / 42%)
• Train: 4,000 (80 shards) · Val: 500 (10 shards) · Test: 500 (10 shards)

What changed from v1

axis	v1 (skyjepa-quadrotor-sim)	v2 (this dataset)
airframe	±50% mass, ±30% inertia, ±50% thrust/torque	narrow: ±12% mass, ±10% inertia, ±12% thrust/torque, motor τ∈[0.02,0.06] s, drag∈[0.15,0.35]
references	100% aggressive GP paths	~60% mission-shaped (waypoint legs + hovers + vertical legs) / ~40% aggressive GP
wind	none	steady vector (horizontal U[0,8] m/s, random heading; vertical U[−0.5,0.5]) + Ornstein–Uhlenbeck gusts (σ≈1 m/s, τ≈1 s), entered as air-relative drag
payload	none	scheduled mid-flight mass steps (drop 0.3→0 or attach 0→0.3 kg); at a hover for mission legs (the Release), mid-flight for aggressive
air density	none	per-trajectory factor U[0.72, 1.0] scaling thrust & drag (sea level → ~3000 m)
battery sag	none	thrust decays linearly over the flight by U[0, 12%]

Wind is not in the 18-dim observed state — the latent must infer it (a future wind-readout head depends on this). Collection controllers are privileged (true airframe params) but wind-blind, so feedback — not feedforward — fights the wind. After a payload step the controller is told the new mass only after a ~0.5 s reaction delay, so the transient is in the data while the rollout still recovers.

Schema

Each split is a set of compressed .npz shards named {split}_{id}.npz:

array	shape	dtype	meaning
states	(n, 201, 18)	float32	18-dim state: position(3), velocity(3), rotation matrix(9), body rate(3)
actions	(n, 200, 4)	float32	4-dim action per step (motor commands)
eta	(n, 22)	float32	per-trajectory parameters: 13 airframe dims + 9 disturbance dims (wind xyz, gust σ, density ρ, sag, initial payload, event time, event Δ)
is_mission	(n,)	bool	mission-shaped (True) vs aggressive GP (False)
attempt_*	(a,)	—	per-attempt wind magnitude / rejected / mission flags (incl. diverged rollouts) for auditing divergence-rejection by wind band

There are N states and N-1 actions per trajectory (20 Hz, 10 s). Splits are disjoint at the shard level (deterministic 80/10/10 by shard id). The window loader (TrajectoryWindows) reads states/actions/eta and ignores the rest, so it works unchanged.

Usage

pip install "skyjepa[hf] @ git+https://github.com/edgarmoreaualix/LeDrone.git#subdirectory=skyjepa"
python scripts/download_dataset_hf.py --repo-id Ethgar/skyjepa-mission-5k --out-dir data/mission5k
python scripts/run_train.py --profile full --stage both --data-dir data/mission5k

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.