On Surprising Effects of Risk-Aware Domain Randomization for Contact-Rich Sampling-based Predictive Control

In a new preprint, researchers from Caltech (Sergio Esteban, Junheng Li, Vince Kurtz, and Aaron Ames) explore a previously overlooked application of domain randomization (DR) in contact-rich sampling-based predictive control (SPC). While DR is widely used in policy learning to train models that generalize across simulated variations, it has rarely been applied directly inside an SPC loop, where rollout quality is highly sensitive to uncertainty. The team tests their risk-aware DR approach on the Push-T task—a simple but representative planar pushing scenario—by evaluating three rollout aggregation strategies: average (mixing all randomizations equally), optimistic (using the best outcome), and pessimistic (focusing on worst-case).

The results reveal a surprising dual effect: DR not only makes the controller more robust to model mismatch, but also systematically reshapes the effective cost landscape seen by the sampling optimizer. Specifically, the basin of attraction around contact-producing actions (e.g., pushing or grasping) is widened when using optimistic aggregation, making the controller more likely to engage in contact-rich behaviors even under significant model uncertainty. This opens the door to designing better-grounded risk-aware SPC methods for real-world robotic tasks where contacts are inevitable and model errors abound. The paper includes 3 figures and a video demonstration.