This was an old idea to improve IMU-only dead-reckoning. Preintegration is a well-known technique that isolates a state change over an arbitary long duration into state-independent "relative motion increment" (RMI) terms. In theory, these state-independent terms should be learnable as a function of the IMU measurements only.
Hence we train a model to learn the RMIs. However, it seems that a simple affine model
is able to achieve 99% of the improvement in terms of error reduction. More sophisticated models have largely diminishing returns.