Reproduce “Object-centric proto-symbolic behavioural reasoning from pixels” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruben van Bergen, Justus Hübotter, Alma Lago, Pablo Lanillos
The Object-centric Behavioural Reasoner (OBR), a brain-inspired deep learning architecture, learns to interpret, control, and reason about its environment from pixel-level input by leveraging object-centric representations. This system demonstrated emergent conditional behavioral reasoning, generalization across varying object counts, and online adaptation to environmental changes in synthetic 2D and 3D scenes.
Key insights
- Object-centric representations, learned unsupervised from pixel data and grounded in action, provide an effective interface for integrating perception, control, and emergent proto-symbolic reasoning.
- Iterative inference mechanisms in the perceptual backbone enhance object shape fidelity and facilitate online adaptation to dynamic scene changes, contributing to robust environmental interaction.
- A modular design, combining object-centric perception, a lightweight linear state-space dynamics model, and a self-supervised preference network, enables complex conditional behavioral reasoning and logical composition without explicit, rule-specific training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.