DreamSAC: Learning Hamiltonian World Models via Symmetry Exploration

TL;DR

DreamSAC introduces a Hamiltonian-based curiosity-driven exploration method that enables learning physical invariances, leading to improved extrapolative generalization in 3D physics simulations.

Contribution

It proposes a novel symmetry exploration strategy and a Hamiltonian-based world model with a contrastive learning objective for better physical invariance learning.

Findings

Outperforms state-of-the-art baselines in extrapolative 3D physics tasks

Effectively learns physical invariances from raw pixel data

Enhances robustness of world models in novel environments

Abstract

Learned world models excel at interpolative generalization but fail at extrapolative generalization to novel physical properties. This limitation arises because they learn statistical correlations rather than the environment's underlying generative rules, such as physical invariances and conservation laws. We argue that learning these invariances is key to robust extrapolation. To achieve this, we first introduce \textbf{Symmetry Exploration}, an unsupervised exploration strategy where an agent is intrinsically motivated by a Hamiltonian-based curiosity bonus to actively probe and challenge its understanding of conservation laws, thereby collecting physically informative data. Second, we design a Hamiltonian-based world model that learns from the collected data, using a novel self-supervised contrastive objective to identify the invariant physical state from raw, view-dependent pixel observations. Our framework, \textbf{DreamSAC}, trained on this actively curated data, significantly outperforms state-of-the-art baselines in 3D physics simulations on tasks requiring extrapolation.