Ethgar/fixed-wing-taskfree-1k

Fixed-wing task-free flights (1000 x 30 s)

The data the fixed-wing world model learned from: 1000 simulated flights of 30 s at 20 Hz (8.3 hours), each on its own randomized airframe in its own wind. There is no task, no reward and no goal, only the machine.

• Airframe randomization per flight: mass ±15 %, inertia ±20 %, every aerodynamic coefficient ±15 %, motor ±15 %, steady wind up to 3 m/s in a random direction, plus Ornstein-Uhlenbeck gusts. Servo (40 ms) and motor (80 ms) lags.
• Flown by a simple autopilot following smooth random references (bank ±40 deg, altitude 15-90 m, airspeed 11-20 m/s), with exploration on its commands. Even flights (kind = 0) get smooth exploration (0.5 s correlation). Odd flights (kind = 1) get white noise every step, the jerky sequences a sampling planner tries.
• 1000 flights, none crashed. Generated on a laptop in 37 minutes (datagen.py, seed 0).

Files

data/shard_000.npz ... shard_019.npz, 50 flights each:

key	shape	content
states	(50, 600, 18) float32	position (m, x east / y north / z up), velocity (m/s, world), rotation matrix columns (body x forward / y left / z up -> world), body rates (rad/s)
actions	(50, 600, 4) float32	commanded throttle (0-1), aileron, elevator, rudder (rad, ±0.35)
eta	(50, 40) float32	the flight's airframe: mass (1), inertia diagonal (3), aero coefficients sorted by name (32), motor constant (1), steady wind (3)
kind	(50,) int	0 smooth exploration, 1 step-wise (MPPI-like)

states[:, k] is measured before actions[:, k] is applied for 50 ms.

By Edgar Moreau · huggingface.co/Ethgar