A Mechanistic Analysis of Sim-and-Real Co-Training in Generative Robot Policies

arXiv cs.RO / 4/16/2026

💬 OpinionSignals & Early TrendsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper studies why sim-and-real co-training works for generative robot policies, addressing a gap in understanding the underlying mechanisms despite its empirical success.
It identifies two intrinsic effects that govern performance: structured representation alignment as the primary driver and an importance reweighting effect as a secondary modifier.
The structured representation alignment effect captures a trade-off between aligning representations across domains and maintaining enough domain discernibility for robust policy learning.
The importance reweighting effect is attributed to domain-dependent modulation of action weighting during training.
The authors validate these claims through both controlled toy-model experiments and extensive sim-and-sim as well as sim-and-real robot manipulation experiments, and propose a simple method that improves on prior approaches.

Abstract

Co-training, which combines limited in-domain real-world data with abundant surrogate data such as simulation or cross-embodiment robot data, is widely used for training generative robot policies. Despite its empirical success, the mechanisms that determine when and why co-training is effective remain poorly understood. We investigate the mechanism of sim-and-real co-training through theoretical analysis and empirical study, and identify two intrinsic effects governing performance. The first, \textbf{``structured representation alignment"}, reflects a balance between cross-domain representation alignment and domain discernibility, and plays a primary role in downstream performance. The second, the \textbf{``importance reweighting effect"}, arises from domain-dependent modulation of action weighting and operates at a secondary level. We validate these effects with controlled experiments on a toy model and extensive sim-and-sim and sim-and-real robot manipulation experiments. Our analysis offers a unified interpretation of recent co-training techniques and motivates a simple method that consistently improves upon prior approaches. More broadly, our aim is to examine the inner workings of co-training and to facilitate research in this direction.