Running a Real Policy#
The zero-action experiments keep the robot still and success rates at zero. In this section we will see actual pre-trained models in action. Arena ships clients for two foundation-model policy servers. On the Rubik’s-cube pick-and-place task used below, openpi’s pi05 generally lands non-zero success rates zero-shot, while GR00T N1.6-DROID gets close to zero on most object variants; start with openpi for a more interactive first run, and try GR00T for a contrasting baseline.