mintar
commented
2 years ago Almost all of that noise does not come from the gaussianNoise parameter, but is unintentional. Here's a plot of your experiment (rotating robot) where I set gaussianNoise = 0:
It looks almost identical to your plot. The noise is the same when the robot is stationary. I think this is because even when the robot is stationary in Gazebo, there are always small corrections being applied. There are a lot of simulated forces at work: there is a repulsive force between the ground and each of the wheels; the powered wheels are being commanded to turn forwards and backwards to keep them still, and so on. All of this leads to the robot to "vibrate" by a minuscule amount. Unfortunately, the gazebo_ros_imu_sensor implementation just reads the instantaneous linear accelerations and angular velocities and returns those. What it should do is calculate the mean for each sampling period (the plugin is updated at 1000 Hz, but only publishes at 50 Hz).
Apart from that, the gazebo_ros_imu_sensor implementation of noise is just plain horrible. There's only one parameter (gaussianNoise) that's applied equally to orientation (as a unitless quaternion), linear acceleration (in m/s²) and angular velocity (in rad/s). Instead, it should at least have these parameters:
- gyroscope white noise N_g, rad/√s (noise density)
- acceleration white noise N_a, (m/s)/√s (noise density)
... and possibly also:
- gyroscope random walk K_g, (rad/s)/√s (random walk)
- acceleration random walk K_a, (m/s²)/√s (random walk)
Then the total noise can be properly calculated like this:
import numpy as np
# n: number of measurements
# dt: sampling rate [s], e.g. dt = 1e-2 when sampling at 100 Hz
def generate_signal(n, dt, noise_density, random_walk, random_state=0):
rng = np.random.RandomState(random_state)
white = noise_density / np.sqrt(dt) * rng.randn(n)
walk = random_walk * np.sqrt(dt) * np.cumsum(rng.randn(n))
return white + walkThe orientation should probably be either not filled at all or be returned without noise, because in order to be realistic, it should be coupled to the gyroscope and accelerometer noise. There are already packages like imu_filter_madgwick that compute the orientation from gyro and accelerometer data, but it doesn't make sense to replicate that here.
So in short: If you want to get rid of the noise, you'll have to improve gazebo_ros_imu_sensor!
narutojxl
Hello authors,
Thanks for your sharing this interesting software. I'm using your mir_robot noetic branch. When remaining the default imu noise params
<xacro:property name="imu_stdev" value="0.00017" />ofmir_100_v1.urdf.xacro, which loadsimu.gazebo.urdf.xacroand sets<xacro:imu_gazebo link="${prefix}imu_link" imu_topic="/imu/data" update_rate="50.0" stdev="${imu_stdev}" />. imu's gaussianNoise =0.00017 * 0.00017 is very very small.This is
imu.gazebo.urdf.xacrocontent:rostopic pub /cmd_vel geometry_msgs/Twist -r 3 -- '[0,0,0]' '[0, 0, 0.4]'to let mir rotate in place. I found the imu's acc and gyro_z are very noisy. First picture is acc_x and acc_y, second is acc_z, third is gyro_x and gyro_y, the last one is gyro_z. For example, acc_x often is 2 m/s^2, acc_y often is 1m/s^2, acc_z often is 8.5 to 11 m/s^2. Command angular z velocity is 0.4 rad/s, but the gyro_z is 0.36 rad/s to 0.42 rad/s.