I wrote a camera/IMU tracking software (VIO) before. I can confirm this statement. Get it working is difficult and requires getting all the algos right. Getting it work reliably is a lot of rigorous testing, collecting datasets, evaluating slightest changes, tuning numerical stability, experimenting with different techniques which might work better in low light and optimizing for the hardware (VIO is intensive but the best VIO is useless if there is no computation power left for the app itself).

Neither of this is required for a slick demo in well lit room and perhaps on better HW.