Geometric Red-Teaming: Auto-Generating Object Variants That Break Robot Policies

Forum topic · 二一 · 2026-05-13

Summary

A CoRL 2025 Oral paper introduces Geometric Red-Teaming, an automated method for stress-testing robot manipulation policies. Instead of manually designed adversarial scenarios, the system takes a target object and a trained grasp/insertion policy, then automatically generates 'CrashShapes'—physically plausible geometric deformations that still look like the original object but cause the policy to fail catastrophically, with success rates dropping from 90% to 22.5%. Technically, the method uses Jacobian fields for local deformation combined with gradient-free optimization in simulation to search for failure-prone shapes. Discovered failure modes include subtle shifts of visual features in insertion tasks and tiny changes to contact surfaces in grasping. Beyond red-teaming, the paper demonstrates Blue-Teaming: fine-tuning policies on CrashShapes restores success rates to around 90% on those deformed variants while preserving performance on original objects.

Testing robot policy robustness has traditionally relied on manually designed scenarios—but the set of adversarial cases a human can imagine is limited. Geometric Red-Teaming offers an automated alternative.

The Method

This CoRL 2025 Oral paper proposes the following: given a target object (e.g., a mechanical part) and a trained grasping policy, the system automatically generates CrashShapes—geometric deformations of the object that remain physically plausible (they still look like the original part) but cause the robot policy to collapse completely, with success rates falling from 90% to 22.5%.

How It Works

Local deformations are computed using a Jacobian field
Gradient-free optimization searches in simulation for the shapes most likely to make the policy fail
Discovered failure modes include:
Subtle shifts of visual features in insertion tasks
Tiny changes to contact surfaces in grasping tasks

Beyond Red-Teaming: Blue-Teaming

This isn't just a red-team tool. The paper also demonstrates Blue-Teaming: fine-tuning the policy on the discovered CrashShapes restores success rates to ~90% on those deformed shapes, while maintaining performance on the original objects.

Takeaway

The best defense isn't building the highest wall—it's knowing where the enemy will attack.

Source: [Geometric Red-Teaming for Robotic Manipulation / CoRL 2025 Oral]