Robustness

A system's ability to keep performing without much degradation when inputs or the environment carry noise or small changes.

Robustness means a system's ability to keep working properly when inputs contain errors, the environment is perturbed, or conditions change. It overlaps with generalization but emphasizes something different: generalization asks “does it still work with a new object or a new scene,” while robustness is more concerned with whether the same task falls apart under disturbances such as lighting changes, a shifted camera, sensor noise, or being bumped by a person. Real-world disturbances are everywhere, so robots are held to a high standard here. LIBERO-Plus (2025) systematically adds seven categories of perturbation to the LIBERO simulation benchmark and finds that some VLA models' success rates drop from 95% to under 30% with only a slight change in camera viewpoint or initial state, and that the models often ignore the language instruction as well. Common ways to improve robustness include domain randomization, data augmentation, and training with deliberate perturbations.

ExampleA quadruped robot gets kicked from the side while walking but adjusts its steps and stays standing; a robot-arm policy still manages to grasp an object even after its camera has been knocked a few centimeters out of place.

Related

Generalization · Domain Randomization · Data Augmentation · Out-of-Distribution · Generalization / Robustness Evaluation · Push Recovery

Sources

Robustness (computer science) - Wikipedia
LIBERO-Plus: In-depth Robustness Analysis of Vision-Language-Action Models

As of

2025-10