Object Representations for Learning and Reasoning

Thirty-fourth Conference on Neural Information Processing Systems (NeurIPS)

December 11, 2020, Virtual Workshop

Join via the livestream 🎥 · RocketChat · @ORLR_Workshop · #ORLR2020 · Join our community Slack!

OGRE: An Object-based Generalization for Reasoning Environment

  • Kelsey Allen, Anton Bakhtin, Kevin Smith, Joshua Tenenbaum, and Laurens van der Maaten
  • (Oral)
  • PDF


If an agent understands how to reason about some objects, can it generalize this understanding to new objects that it has never seen before? We propose the Object-based Generalization for Reasoning Environment (OGRE) for testing object generalization in the context of active physical reasoning. OGRE emphasizes evaluating agents by how efficiently they solve novel physical reasoning tasks, not just how well they can predict the future. OGRE provides two levels of generalization: generalization over reasoning strategies with familiar objects, and generalization over new object types that still share similar material properties to those in training. We run three baseline agents on OGRE, showing that an image-based Deep Q-Network can learn reasoning strategies that generalize in a limited way across familiar object types, but does not generalize at all to new object types. We hope OGRE will encourage advances in building object representations that more explicitly enable generalizable reasoning and planning compared to previous benchmarks.