Mehsoft & Korvax
Mehsoft Mehsoft
Korvax Korvax
Mehsoft Mehsoft
Hey, I've been tweaking the sensor fusion algorithm for our drone fleet. Have you tested the new Kalman filter implementation?
Korvax Korvax
The Kalman filter's convergence rate increased by 12%, but I noticed a persistent bias in the velocity estimate; I’d recommend re‑tuning the process noise covariance for the high‑roll scenarios. Also, the residuals are still not Gaussian—perhaps add a robust outlier rejection step before the update.
Mehsoft Mehsoft
Good data, but that bias is a flag. I’ll tweak Q for roll‑heavy dynamics and drop a Huber loss before the update. Let’s see if the residuals calm down.
Korvax Korvax
Great plan, but keep an eye on the numerical stability of the covariance update when you drop the Huber loss—numerical conditioning can get hairy. Once you adjust Q, we should run a quick Monte Carlo to verify the residual distribution.
Mehsoft Mehsoft
Got it, will guard against covariance blow‑up with a condition‑number check and a fallback to the classic update if it exceeds a threshold. Once Q is tuned I’ll script a 1k‑run Monte Carlo and plot the residual histogram. This should finally make the residuals look Gaussian, or at least not like a bad data set.
Korvax Korvax
Excellent. Make sure the fallback preserves symmetry in the covariance update; any asymmetry will propagate downstream. When you plot, also overlay the theoretical Gaussian PDF so we can quantify the Kolmogorov‑Smirnov distance. That will give us a solid metric to decide if we’re done.