ENH: seed sensor measurement noise per instance

rocketpy/sensors/accelerometer.py

@@ -78,6 +78,7 @@ def __init__(
         cross_axis_sensitivity=0,
         consider_gravity=False,
         name="Accelerometer",
+        seed=None,
     ):
         """
         Initialize the accelerometer sensor
@@ -193,6 +194,7 @@ def __init__(
             temperature_scale_factor=temperature_scale_factor,
             cross_axis_sensitivity=cross_axis_sensitivity,
             name=name,
+            seed=seed,
         )
         self.consider_gravity = consider_gravity
         self.prints = _InertialSensorPrints(self)

rocketpy/sensors/barometer.py

@@ -62,6 +62,7 @@ def __init__(
         temperature_bias=0,
         temperature_scale_factor=0,
         name="Barometer",
+        seed=None,
     ):
         """
         Initialize the barometer sensor
@@ -132,6 +133,7 @@ def __init__(
             temperature_bias=temperature_bias,
             temperature_scale_factor=temperature_scale_factor,
             name=name,
+            seed=seed,
         )
         self.prints = _SensorPrints(self)
 

rocketpy/sensors/gnss_receiver.py

@@ -1,5 +1,3 @@
-import numpy as np
-
 from ..mathutils.vector_matrix import Matrix, Vector
 from ..prints.sensors_prints import _GnssReceiverPrints
 from .sensor import ScalarSensor
@@ -40,6 +38,7 @@ def __init__(
         altitude_accuracy=0,
         velocity_accuracy=0,
         name="GnssReceiver",
+        seed=None,
     ):
         """Initialize the Gnss Receiver sensor.
 
@@ -59,7 +58,7 @@ def __init__(
         name : str
             The name of the sensor. Default is "GnssReceiver".
         """
-        super().__init__(sampling_rate=sampling_rate, name=name)
+        super().__init__(sampling_rate=sampling_rate, name=name, seed=seed)
         self.position_accuracy = position_accuracy
         self.altitude_accuracy = altitude_accuracy
         self.velocity_accuracy = velocity_accuracy
@@ -96,12 +95,12 @@ def measure(self, time, **kwargs):
         ) + Vector(u[3:6])
 
         # Apply accuracy to the position
-        x = np.random.normal(x, self.position_accuracy)
-        y = np.random.normal(y, self.position_accuracy)
-        z = np.random.normal(z, self.altitude_accuracy)
-        vx = np.random.normal(vx, self.velocity_accuracy)
-        vy = np.random.normal(vy, self.velocity_accuracy)
-        vz = np.random.normal(vz, self.velocity_accuracy)
+        x = self._rng.normal(x, self.position_accuracy)
+        y = self._rng.normal(y, self.position_accuracy)
+        z = self._rng.normal(z, self.altitude_accuracy)
+        vx = self._rng.normal(vx, self.velocity_accuracy)
+        vy = self._rng.normal(vy, self.velocity_accuracy)
+        vz = self._rng.normal(vz, self.velocity_accuracy)
 
         self.measurement = (x, y, z, vx, vy, vz)
         self._save_data((time, *self.measurement))

rocketpy/sensors/gyroscope.py

@@ -78,6 +78,7 @@ def __init__(
         cross_axis_sensitivity=0,
         acceleration_sensitivity=0,
         name="Gyroscope",
+        seed=None,
     ):
         """
         Initialize the gyroscope sensor
@@ -193,6 +194,7 @@ def __init__(
             temperature_scale_factor=temperature_scale_factor,
             cross_axis_sensitivity=cross_axis_sensitivity,
             name=name,
+            seed=seed,
         )
         self.acceleration_sensitivity = self._vectorize_input(
             acceleration_sensitivity, "acceleration_sensitivity"

rocketpy/sensors/sensor.py

@@ -59,6 +59,7 @@ def __init__(
         temperature_bias=0,
         temperature_scale_factor=0,
         name="Sensor",
+        seed=None,
     ):
         """
         Initialize the accelerometer sensor
@@ -141,6 +142,12 @@ def __init__(
         self._save_data = self._save_data_single
         self._random_walk_drift = 0
         self.normal_vector = Vector([0, 0, 0])
+        # Per-instance RNG seeded deterministically (from the scenario seed when
+        # provided) so sensor noise is reproducible and independent of the
+        # process-global numpy RNG -- and therefore safe under parallel/forked
+        # evaluation. ``seed=None`` keeps the noise random but still per-instance.
+        self._seed = seed
+        self._rng = np.random.default_rng(seed)
 
         # handle measurement range
         if isinstance(measurement_range, (tuple, list)):
@@ -353,6 +360,7 @@ def __init__(
         temperature_scale_factor=0,
         cross_axis_sensitivity=0,
         name="Sensor",
+        seed=None,
     ):
         """
         Initialize the accelerometer sensor
@@ -469,6 +477,7 @@ def __init__(
                 temperature_scale_factor, "temperature_scale_factor"
             ),
             name=name,
+            seed=seed,
         )
 
         self.orientation = orientation
@@ -553,12 +562,12 @@ def apply_noise(self, value):
         """
         # white noise
         white_noise = Vector(
-            [np.random.normal(0, self.noise_variance[i] ** 0.5) for i in range(3)]
+            [self._rng.normal(0, self.noise_variance[i] ** 0.5) for i in range(3)]
         ) & (self.noise_density * self.sampling_rate**0.5)
 
         # random walk
         self._random_walk_drift = self._random_walk_drift + Vector(
-            [np.random.normal(0, self.random_walk_variance[i] ** 0.5) for i in range(3)]
+            [self._rng.normal(0, self.random_walk_variance[i] ** 0.5) for i in range(3)]
         ) & (self.random_walk_density / self.sampling_rate**0.5)
 
         # add noise
@@ -655,6 +664,7 @@ def __init__(
         temperature_bias=0,
         temperature_scale_factor=0,
         name="Sensor",
+        seed=None,
     ):
         """
         Initialize the accelerometer sensor
@@ -726,6 +736,7 @@ def __init__(
             temperature_bias=temperature_bias,
             temperature_scale_factor=temperature_scale_factor,
             name=name,
+            seed=seed,
         )
 
     def quantize(self, value):
@@ -763,15 +774,15 @@ def apply_noise(self, value):
         """
         # white noise
         white_noise = (
-            np.random.normal(0, self.noise_variance**0.5)
+            self._rng.normal(0, self.noise_variance**0.5)
             * self.noise_density
             * self.sampling_rate**0.5
         )
 
         # random walk
         self._random_walk_drift = (
             self._random_walk_drift
-            + np.random.normal(0, self.random_walk_variance**0.5)
+            + self._rng.normal(0, self.random_walk_variance**0.5)
             * self.random_walk_density
             / self.sampling_rate**0.5
         )

tests/fixtures/sensors/sensors_fixtures.py

@@ -62,6 +62,7 @@ def noisy_barometer():
         operating_temperature=25 + 273.15,
         temperature_bias=0.02,
         temperature_scale_factor=0.02,
+        seed=42,
     )
 
 

tests/unit/sensors/test_sensor_seeding.py

@@ -0,0 +1,99 @@
+"""Determinism tests for seeded sensor noise (additive, isolated).
+
+Sensor noise is drawn from a per-instance ``numpy.random.Generator`` seeded
+deterministically, instead of the process-global ``numpy.random``. This makes a
+seed-reproducible run (e.g. a judged competition) yield the identical sensor
+noise for a given seed, regardless of the global RNG state or parallel/forked
+execution -- mirroring the seeding the Monte Carlo layer already uses
+(``stochastic_model.py`` / ``monte_carlo.py``).
+
+The tests construct sensors directly and do not touch the existing fixtures or
+inherited tests. Sensor noise is sampled on a fixed time grid (not per adaptive
+solver step), so a fixed number of sequential draws is a faithful stand-in for a
+flight of a given duration.
+"""
+
+import numpy as np
+
+from rocketpy.mathutils.vector_matrix import Vector
+from rocketpy.sensors.accelerometer import Accelerometer
+from rocketpy.sensors.gnss_receiver import GnssReceiver
+
+
+def _accelerometer(seed):
+    # Non-zero white noise + random walk so the draws actually exercise the RNG.
+    return Accelerometer(
+        sampling_rate=10,
+        noise_density=1.0,
+        noise_variance=1.0,
+        random_walk_density=0.5,
+        random_walk_variance=1.0,
+        seed=seed,
+    )
+
+
+def _noise_sequence(sensor, n=16):
+    return [tuple(sensor.apply_noise(Vector([0.0, 0.0, 0.0]))) for _ in range(n)]
+
+
+def test_same_seed_is_reproducible():
+    assert _noise_sequence(_accelerometer(42)) == _noise_sequence(_accelerometer(42))
+
+
+def test_different_seeds_decorrelate():
+    assert _noise_sequence(_accelerometer(1)) != _noise_sequence(_accelerometer(2))
+
+
+def test_noise_independent_of_global_numpy_rng():
+    # Perturbing the process-global RNG must NOT change a seeded sensor's noise.
+    # This is the regression guard for the original bug (noise drawn from the
+    # global ``np.random``).
+    np.random.seed(0)
+    first = _noise_sequence(_accelerometer(7))
+    np.random.seed(999)
+    _ = [np.random.random() for _ in range(1000)]
+    second = _noise_sequence(_accelerometer(7))
+    assert first == second
+
+
+def test_seeded_sensor_does_not_consume_global_rng():
+    # A seeded sensor draws only from its own generator, leaving the global RNG
+    # position untouched -- so it cannot contaminate other code or parallel envs.
+    np.random.seed(0)
+    position_before = np.random.get_state()[2]
+    _noise_sequence(_accelerometer(7))
+    position_after = np.random.get_state()[2]
+    assert position_before == position_after
+
+
+def test_recreating_generator_from_seed_reproduces_sequence():
+    # Reproducibility comes from the seed: re-creating the generator from the
+    # same seed (which is what a freshly-constructed sensor does) replays the
+    # same sequence. The random walk is cumulative, so zeroing it matters too.
+    sensor = _accelerometer(99)
+    first = _noise_sequence(sensor)
+    sensor._rng = np.random.default_rng(sensor._seed)
+    sensor._random_walk_drift = Vector([0.0, 0.0, 0.0])
+    assert _noise_sequence(sensor) == first
+
+
+def _gnss_sequence(seed, n=8):
+    gnss = GnssReceiver(
+        sampling_rate=1,
+        position_accuracy=5.0,
+        altitude_accuracy=5.0,
+        velocity_accuracy=1.0,
+        seed=seed,
+    )
+    # Minimal launch-frame state: 100 m up, identity attitude quaternion.
+    state = np.array([0, 0, 100, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0], dtype=float)
+    measurements = []
+    for _ in range(n):
+        gnss.measure(0.0, u=state, relative_position=Vector([0.0, 0.0, 0.0]))
+        measurements.append(gnss.measurement)
+    return measurements
+
+
+def test_gnss_is_seeded_and_reproducible():
+    assert _gnss_sequence(5) == _gnss_sequence(5)
+    assert _gnss_sequence(5) != _gnss_sequence(6)