Reproduce “Fundamental Challenges in Deep Learning for Stiff Contact Dynamics” on OpenResearch

About this paper

Mihir Parmar, Mathew Halm, Michael Posa

Deep learning models exhibit substantial performance degradation, including reduced training efficiency and persistent generalization errors, when attempting to learn the non-smooth dynamics of stiff frictional contact. This effect is largely due to the models' inherent inductive bias towards smooth interpolations, posing challenges for real-world robotics and simulation realism.

Key insights

• The smooth-interpolating bias of standard deep learning methods fundamentally conflicts with the non-smooth dynamics of stiff frictional contact, limiting their efficacy.
• Generalization error for stiff contact models does not diminish with increased training data, challenging common deep learning assumptions.
• The use of unrealistically soft contact settings in simulations can mask true performance limitations, contributing significantly to the sim-to-real gap.

Summary by alphaXiv.