Reproduce “What Do They See? Interpreting Complex Road Scenarios Through the Eyes of Vision-Language-Action Models for Safe and Trustworthy Autonomous Vehicle Learning” 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

Kalpana Panda, Wesley Maia, Vinti Agarwal, Ross Greer

Researchers from Birla Institute of Technology and Science, Pilani and University of California, Merced developed the Counterfactual Vision Action Analysis (CVAA) framework to interpret how vision-language-action (VLA) models for autonomous driving perceive road scenarios. This framework, using photorealistic object removal, revealed that VLA models often do not attribute causal influence to objects in a human-intuitive way and exhibit complex internal processing pathways.

Key insights

  • VLA models do not consistently perceive and weigh scene objects in a human-intuitive manner, with object importance being unstable and sometimes improving ground-truth accuracy upon removal.
  • Causal influence propagates through complex internal pathways, with four distinct regimes identified (coupled, transparent, decoupled, silent variants), and the trajectory expert acts as a sensitive nonlinear amplifier.
  • Standard language model interpretability tools (like Logit Lens) are partially inadequate for VLA architectures, as visual tokens encode scene geometry and relational context rather than discrete categorical object labels.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.