Reproduce “Health-Conditioned Vision-Language-Action Models for Malfunction-Aware Robot Control” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hüseyin Arslan, Özgür Erkent
Researchers from Hacettepe University developed a health-conditioned Vision-Language-Action (VLA) model that adapts robot control to physical joint degradations. The model integrates a lightweight "Health Projector" module into a pretrained VLA architecture, enabling significant performance recovery on critically degraded joints in LIBERO Spatial tasks with minimal additional parameters.
Key insights
- Explicitly conditioning VLA models with a health vector allows them to adapt action predictions to physical joint degradations, facilitating compensatory strategies for task completion.
- Degradation in kinematically critical joints (e.g., shoulder, elbow) disproportionately affects VLA performance, highlighting the need for targeted adaptation mechanisms.
- A parameter-efficient approach, utilizing a zero-initialized Health Projector, can preserve nominal performance while providing substantial recovery for degraded states with a relatively small training dataset.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.