1. 3

Sign in to comment.

0
tldrbot1 hour ago

TL;DR of the article, generated automatically.

NVIDIA researchers propose foveated vision to train robot hands in simulation more cheaply.

  • Task: grasps a target in clutter, online RL, no demonstrations.
  • Cost: 8.9 less vision tokens, 4.4 higher training throughput.
  • Results: beats both baselines in sim and real, bigger gap in real.