Question about IMU noise implementation and cov_decoder function

[ProChampion841]

Hi, @haleqiu , @cannnnxu , How are you?

I am Brandon, a researcher at Texas Tech University. I have been studying your AirIMU repository and the paper (https://arxiv.org/html/2310.04874v4), and I have a few technical questions:

1. IMU Noise Implementation
   The paper mentions corrected acceleration and angular velocity with stochastic noise η_acc(k) and η_gyro(k), but the code seems not to add these explicitly. Is the noise modeled implicitly by the network?

2. Determining gyro_std and acc_std
   How should these standard deviations be determined in practice? Should they be computed per sensor, or are there recommended default values?

3. cov_decoder Function Logic
   In cov_decoder, the code uses:

   acc = torch.exp(self.acccov_decoder(x) - 5.)
   gyro = torch.exp(self.gyrocov_decoder(x) - 5.)

   Could you explain why 5 is subtracted before exponentiation?

I am eager to understand these details as I plan to apply AirIMU to my research and hope to collaborate or contribute to the project.

Thank you very much for your time and guidance.
Best regards,
Brandon.

[haleqiu]

@ProChampion841 Hi,

1. The model explicitly model the η_acc(k) and η_gyro(k), I will say it was supervised implicitly
2. the acc std is just a initalization for the correction decoder. If you remove it, it will still converge maybe with a few more iterations.
3. -5 is to make the training starting at a better initial position where we randomly pick previously. If you remove it, it will still converge maybe with a few more iterations. But to be honest will not change too much.