Zero-Shot Sim-to-Real Robot Learning: A Dexterous Manipulation Study on Reactive Catching

Researchers introduced DRIS, a training method that simulates multiple randomized dynamics at once to improve zero-shot sim-to-real learning.

The approach produced robust reactive catching policies for a difficult flat-plate task that transferred directly from simulation to a real robot.

It suggests a practical way to handle real-world uncertainty in dexterous manipulation without costly on-robot fine-tuning, speeding deployment.