Reproduce “Quantifying and Visualizing Sim-to-Real Gaps: Physics-Guided Regularization for Reproducibility” on OpenResearch

About this paper

Yuta Kawachi

DENSO IT Laboratory researchers developed a physics-guided regularization method combined with parameter-conditioned domain randomization to improve sim-to-real transfer for robotic control, particularly on systems with large sim-to-real gaps. The approach demonstrated enhanced transfer fidelity and reduced anomalous oscillations on a high-gear-ratio balancing robot, achieving closer alignment between simulated and real-world angular velocity recovery compared to traditional domain randomization.

Key insights

• Unconstrained neural network policies trained solely with domain randomization in simulation tend to develop unrealistically high equivalent gains, which manifest as unstable oscillations on physical hardware.
• Naïve domain randomization can bias the training process towards 'easy' low-loss regions in the parameter space that may not correspond to the real hardware, leading to degraded performance at nominal plant parameters.
• Integrating lightweight empirical physics guidance (measured proportional gains) into the learning process is effective for explicitly embedding real-world actuation characteristics and closing sim-to-real gaps.

Summary by alphaXiv.