[Paper Review] Relational inductive bias for physical construction in humans and machines

Jessica B. Hamrick, Kelsey R. Allen+5 more|arXiv (Cornell University)|Jun 4, 2018

Design Education and PracticeEngineering33 references62 citations

TL;DR

The paper argues that relational inductive bias—structured object- and relation-centric representations—enables effective physical construction tasks and demonstrates a graph-network-based RL agent that outperforms humans and non-relational baselines on a gluing task.

ABSTRACT

While current deep learning systems excel at tasks such as object classification, language processing, and gameplay, few can construct or modify a complex system such as a tower of blocks. We hypothesize that what these systems lack is a "relational inductive bias": a capacity for reasoning about inter-object relations and making choices over a structured description of a scene. To test this hypothesis, we focus on a task that involves gluing pairs of blocks together to stabilize a tower, and quantify how well humans perform. We then introduce a deep reinforcement learning agent which uses object- and relation-centric scene and policy representations and apply it to the task. Our results show that these structured representations allow the agent to outperform both humans and more naive approaches, suggesting that relational inductive bias is an important component in solving structured reasoning problems and for building more intelligent, flexible machines.

Motivation & Objective

• Motivate the role of relational reasoning in human and machine physical construction abilities.
• Introduce the gluing task to study how relations among blocks influence stability under gravity.
• Develop a graph-network-based deep RL agent with object- and relation-centric representations.
• Empirically compare relational vs. non-relational agents and humans on the gluing task and control tasks.

Proposed method

• Define towers of blocks and a gluing mechanism to stabilize unstable structures under gravity.
• Represent scenes as graphs with nodes (blocks), edges (glue relations), and globals (tower stability) and apply a graph network to encode/predict relations.
• Train RL agents (MLP, GN-FC, GN) and a simulation-based baseline; assess Q-values or action policies from graph emissaries.
• Use supervised experiments to evaluate stability prediction and optimal glue locations under varying tower sizes.
• Compare performance across tower sizes and assess generalization to unseen block counts.

Experimental results

Research questions

• RQ1 Does a relational inductive bias enable better performance on physical construction tasks compared to non-relational approaches?
• RQ2 How does a graph-network based policy compare to humans and to model-free baselines on the gluing task?
• RQ3 What is the impact of explicit physical reasoning (simulation) versus purely relational reasoning on task performance?
• RQ4 Can relational structures generalize to larger or unseen tower sizes beyond training?
• RQ5 What behavioral patterns and strategies do humans exhibit, and how do they differ from GN-based agents?

Key findings

• A relational inductive bias via graph networks significantly improves performance over MLP and GN-FC baselines.
• The GN agent with correct relational structure achieves higher rewards and better generalization than the fully-connected GN-FC agent.
• A simulation-based agent with explicit physics knowledge performs best but GN-based agents still outperform non-relational baselines, indicating relational structure is crucial.
• Humans perform above chance and show top-to-bottom strategies, with slower first glue decisions; GN agents exhibit fewer invalid actions compared to humans.
• Generalization: GN agents show robust performance on unseen tower sizes (7 and 10 blocks) with minimal degradation, while GN-FC struggles to extrapolate to larger towers; MLP fails to generalize across sizes.