Robots and machine learning

What it changed, and did not.

Where it helps

Perception improved enormously: recognising objects, estimating pose, reading scenes. Learned control policies also handle variation better than hand-tuned rules in some tasks.

Training in simulation is cheap; transferring to reality is the hard part.
A physical failure costs hardware, unlike a wrong answer in software.

Where it does not

Data is expensive to collect on real hardware, and guarantees matter when a machine can injure someone. Classical control remains dominant where safety must be provable.