Efficient Use of Incorrectly Parameterized Simulators for Robust Real-World Performance

Develop methods that enable robots to efficiently use simulators with incorrect physical parameterization to learn either a control policy or a stochastic world model that achieves robust real-world performance upon deployment.

Background

The authors argue that minimizing the sim-to-real performance gap is often more practical than eliminating the reality gap, suggesting a focus on learning models that improve control near high-return regions rather than uniformly accurate physics. Prior work shows that optimizing controllers for performance under uncertain parameters can outperform planning with most-likely estimates.

This motivates an open question about principled strategies for exploiting imperfect simulators to learn policies or world models that remain robust when transferred to real hardware.