I'm pretty sure you could use ICA and adaptive subsampling to make a cheap realtime image of your local radio environment with one of these.  Could also a dish antenna on the gymbal to not need ICA.  This would help enough in reusably describing shielded rooms that I'm still trying to move forward on it, after all these years.

The imu is essentially a stream of data.  The timing of this stream can be roughly calibrated by comparing with radio and motor data. (Motor changes, then either radio or imu changes, then the other.)

With the imu seen as a stream, we can constantly slew the motors around regions of interest to acquire data with a high degree of orientation variability.

Here you could train an ML algorithm around the antenna's response, but I'm not used to that for now.

ICA is a form of blind source separation that uses matrices.  For simplification, we bucket the imu values, each one a separate vector of data from the radios.

The radio vectors all have holes in their data, but at first we only care about what the sources are, not what data they are sending, so we store each radio vector as a single FFT, averaged from all recordings in that bucket.

Now the data matrix can be seen as a single audio source, with a huge number of "ears" each receiving a single spectrogram as their "sound".

Blind source separation (ICA, unmixing) remixes the data in such a way that multiply recorded signals cancel each other out, producing as many clear independent signals as their were original recording channels.

After the ICA unmixing, a matrix is produced that holds the weight of every contributing channel.  Because this matrix is the contribution of every angle of recording to every extracted signal, each column shows the spherical responsivity of the antenna and cymbal setup, shifted to be centered around each different source.

The antenna response function of angle could be extracted by aligning and averaging all the columns.  In fact, aligning them is needed to identify the direction of the sources.

There are a _lot_ of unaddressed concerns that will likely require some algorithmic redesign.  It is just one approach.

But it is valuable to have the IMU data be densely available as a stream, if possible.  It is also valuable to support continous rotation drive motors.

Alignment of signals to calculate the antenna response or calibrate the streams could be done by simpy looking for the maximum of the sliding absolute difference or product.

Metrics aiding in the indication of poor data would likely be very helpful.  Approaches like this only work in the regions of data where they are valid.

I did not fully edit this to include further parts.