Is Natural Language Processing a Good Approach to Learning Robotics?

Red: You do see a lot of trends moving towards generalizability, though. Like meta learning, which was a big part of your thesis and all these other setups, like goal-condition learning. For us, like we set up our tasks have a curricula and our simulator has a lot of knobs in it. That makes it possible to make a task just very slightly harder. Even for humans, generalization is hard. It's not as hard as for these systems, but giving it something that's kind of close to what it's good at, but slightly harder. And I'm pretty excited about some of the work and environment design so that you can automatically perturb environments