Yilun Du, MIT: On energy-based models, implicit functions, and modularity

December 21, 2021

RSS · Spotify · Apple Podcasts · Pocket Casts

Yilun Du (Google Scholar) is a graduate student at MIT advised by Professors Leslie Kaelbling, Tomas Lozano-Perez, and Josh Tenenbaum. He's interested in building robots that can understand the world like humans and construct world representations that enable task planning over long horizons.

Some highlights from our conversation

“None of my research is really [about] state-of-the-art. […] The thing that is important to me is that whatever method I come up with, it can do something that prior methods can’t do.”

“If you’re training your agent in a 5-by-5 grid, and then you give it a 10-by-10 grid, it’s never going to generalize. But what if you train the agent on tables in your 5-by-5 grid, right? Like just local tables. Then if I give you a 10-by-10 grid, and I have more tables, you can generalize. So it seems to me like modularity really allows you to generalize in a sense […] even though your global input is completely out-of-distribution, if you process these local modules one by one, it’s much more in-distribution.”

“It feels like we’re not solving the generic robotics problem. You basically train this agent using millions of CPU hours to reorient a single cube in its hand. If I give you a different object, you can’t reorient it. If I put the arm in a different configuration, you can’t reorient it.”

Referenced in this podcast

Thanks to Tessa Hall for editing the podcast.