We were testing our visual SLAM algorithm on robots. We were getting poor performance. Then we calculated wite noise and random walk parameters (using kalibr) for the IMU and used it in our algorithm and the performance included significantly.

Given that calculating random walk requires keeping IMU steady for longer duration (we kept it steady for 20 hours), I am guessing how we can feasibly calculate these values for different deployments. Our robots are meant to operate in broad spectrum of environment (from extreme coldest to extreme hottest etc). Also, we are planning deployment of hundreds of robots in distant future.

Q1. Is it possible to calibrate IMU and obtain its white noise and random walk parameters using some software approach - kinda of "online self calibration" ?

I checked the literature, I came across this paper. It uses transformer based model to predict IMU noise (instead of predicting IMU intrinsics, white noise and random walk). Below are block diagram of the architecture and excerpt describing training strategy.

enter image description here enter image description here enter image description here

Q2. What this paper is exactly doing? Is it reducing error between artificially generated gaussian noise o and noise predicted by model ̂o, that is min L(o, ̂o)? Isnt it just equivalent to train NN to learn gaussian distribution? Isnt any untrained randomly initialied model already mimicking gaussian distribution?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.