Reframing Object-Centric Reinforcement Learning: From Absolute to Relative Representations

Key points

• Object-centric reinforcement learning agents typically receive each object's position in a fixed global frame, even though their decisions usually depend on where objects are relative to one another.1
• In this work, we propose a shift toward Object-Relative Coordinates fixed to Anchor (ORCA), where entities are described in a relative frame of reference.1
• Our results suggest these relative representations improve performance for both MLP and transformer agents.1
• In our controlled Tag environment, we further show that ORCA extends to rotations, that ORBiT generalizes to unseen numbers of distractors, and that these gains require a decision-relevant anchor.1

Sources (1)

• [1]Reframing Object-Centric Reinforcement Learning: From Absolute to Relative Representations
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 20, 01:38 PM

  To operationalize this idea, we introduce Object-Relative inductive-Bias Transformers (ORBiT), a transformer-based architecture for processing anchor-relative inputs in reinforcement learning.

  Object-centric reinforcement learning agents typically receive each object's position in a fixed global frame, even though their decisions usually depend on where objects are relative to one another.

Extractive summary: sentences quoted from the sources.