Paper Reading Group: Towards Practical Multi-object Manipulation Using Relational RL

Towards Practical Multi-object Manipulation using Relational Reinforcement Learning, by Li, Jabri, Darrell, and Agrawal. Made by Andrea Pessl using W&B
Andrea Pessl
The W&B Paper Reading Group will meet next week to discuss Towards Practical Multi-object Manipulation using Relational Reinforcement Learning by Richard Li, Allan Jabri, Trevor Darrell, and Pulkit Agrawal. Learn more and check out some demonstration videos of their simulated robot here.
A live discussion took place on April 26, 10am PT.
The next day, we'll have Professor Agrawal on the W&B Salon to discuss this work, along with his other work on using to deep learning to create robots that can move and act in complex environments (details here).
After you've read the paper, if you have any questions or topics you'd like to discuss in the reading group, you can post them in the comment section below.
Here's the quick summary of the paper:
Reinforcement learning algorithms require an exorbitant number of interactions to learn from sparse rewards. To overcome this, the authors present a simple but effective method for learning from a curriculum of increasing number of objects. They show that attention-based graph neural networks provide critical inductive biases that enable usage of this task curriculum. Their agent achieves a success rate of 75% for stacking 6 blocks, while the existing state-of-the-art method, which uses human demonstrations and resets, only achieves a success rate of 32%.
In this meeting of the paper reading group, W&B deep learning educator Charles Frye will be discussing all about curriculum learning [0], graph neural networks [1, 2], inductive bias [3, 4], & more. We hope to see you there!