IDEA: make a wearable suit with various biometric data collection (motion capture, pressure sensing, EEG & fNIRS, cameras, microphones etc) and collect tons of data,
then train a model to predict the next body state (limb position, voice output, brain state etc) based on new sensory input and previous body state.
If it worked you would have human level virtual agent.

Is this theoretically possible with modern ML?

>>16702633 (OP)

Lotsa unnecessary data for a virtual agent there. More interesting for an exosuit.

Reality: Just do it yourself