Probes and robots are indeed better for exploring space than people. There might even be multiple uses of machine learning in future space exploration. However, it’ll only be used when it is the best option, and not for the sake of it like most “AI” we see today. This is because “AI” is business terminology entirely divorced from the reality of the tech. Machine learning is not genuine intelligence, but a tool for accomplishing goals, and it has very real limitations that make it not useful in all cases.

The main thing that’ll hold back machine learning in probes is power, with probes relying on barely over 100 watts plus whatever dozens of watts their solar panels can generate in good conditions. Current machine learning hardware consumes many watts of energy to run advanced models, so whatever model is used needs to be extremely streamlined and efficient. This power limitation is not related to computing, but nuclear technology and the cost to send matter to space, so unless we pull compact portable fusion out of our asses, probes can only be so smart.