Reproduce “Intervention Design for Effective Sim2Real Transfer” on OpenResearch

About this paper

Melissa Mozifian, Amy Zhang, Joelle Pineau, David Meger

The paper provides a causal inference perspective on sim2real transfer, demonstrating that domain randomization and data augmentation function as interventions to learn invariant latent representations. Its proposed Intervention-based Invariant Transfer learning (IBIT) method enables robust generalization in diverse simulated environments and achieves an 8/10 zero-shot success rate on a real 7DoF robot manipulation task, outperforming baseline approaches.

Key insights

• The effectiveness of domain randomization and data augmentation stems from their role as interventions that help learning algorithms identify and attend to true causal mechanisms invariant to irrelevant visual features.
• Perturbations applied to the environment do not need to be physically realistic; the mere presence of variation along irrelevant dimensions is sufficient for learning invariant features and robust policies.
• Explicitly enforcing invariance through objectives like bisimulation metrics can lead to compact, task-relevant latent representations that generalize robustly across diverse environments.

Summary by alphaXiv.