Health-Conditioned Vision-Language-Action Models for Malfunction-Aware Robot Control

Researchers added a health-conditioning module to a vision-language-action robot policy so it can adapt to joint degradation and other malfunctions during execution.

The method augments the VLA-Adapter with a Health Projector that ingests a robot health vector alongside visual and language inputs.

Trained on 128 teleoperated LIBERO episodes, the model outperformed the baseline on degraded joints, actuator issues, and gripper wear.

The approach offers a lightweight path to more robust robot control in real-world settings where hardware wear and failures are common.