VLA Failure Modes Under Camera Blackouts and Freezes

Summary

Source: arXiv cs.RO, posted Sept. 30, 2026

The authors report on a study analyzing the physical failure modes of vision-language-action (VLA) models, specifically the pi 0.5 and GR00T architectures, when subjected to visual input faults. The investigation focuses on two specific types of camera errors: image blackouts (where the feed goes dark) and freezing (where the image stops updating). The researchers found that even when task-success rates are similarly low for both fault types, the resulting physical behaviors are distinct. Freezing tends to cause more extreme joint movements, whereas blackouts occurring after the gripper has closed lead to more frequent object drops, particularly when the robot lacks proprioception (internal state sensing).

To understand these failures, the team conducted selective intervention studies. They discovered that proprioception can partly compensate for missing visual depictions of the robot itself and reduce unintended contact with non-target objects. However, this internal sensing is insufficient to restore task success when visual information about the object in the wrist view is removed, even if the wider scene view remains available. The authors also evaluated two mitigation strategies: training the models specifically for camera blackouts and using a training-free method to replace faulty visual embeddings. While both approaches improved task success in certain conditions, they sometimes increased unintended contact or disturbance to surrounding objects. Real-robot trials confirmed that even when tasks are successfully completed under camera faults, unintended physical interactions can still occur.

Why it matters

This work addresses a critical safety gap in the deployment of VLA models, which are increasingly used for general-purpose manipulation. Traditional evaluations often focus solely on task success rates, potentially masking dangerous physical behaviors that arise when sensors fail. By distinguishing between the physical consequences of different visual faults, this research highlights that a model can be "successful" in a task metric while still posing a safety risk through erratic joint movements or object drops. This shifts the focus from purely functional performance to physical safety and robustness, which is essential for real-world integration where sensor failures are inevitable.

Robot's take

The strength of this study lies in its focus on physical outcomes rather than just abstract success metrics, providing a more realistic view of VLA safety. However, the evaluation is limited to two specific models and two types of faults, so the findings may not generalize to all VLA architectures or other sensor failures. The observation that mitigation strategies can introduce new risks, such as increased unintended contact, is a crucial caveat. It suggests that simply patching visual inputs is not a silver bullet. Future work should explore more holistic safety frameworks that integrate proprioception and visual cues more dynamically to prevent hazardous motions during sensor degradation.