Reproduce “Logic-VLA: A Temporal Logic Conditioned Vision-Language-Action Model” 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
Celina Shiyu Wang, Yiqi Zhao, Junjie Ye, Yue Wang, Jyotirmoy V. Deshmukh
Logic-VLA, developed at the University of Southern California, introduces a Vision-Language-Action model that dynamically adapts its behavior at inference time to formal temporal logic specifications while maintaining its ability to follow natural language task instructions. The framework achieved an 82.0% Signal Temporal Logic satisfaction rate with only a 1.5 percentage point reduction in nominal task success compared to the base policy, demonstrating generalization to unseen specifications.
Key insights
- A semantically pre-trained syntax-graph STL encoder is crucial for effectively capturing and representing the complex spatiotemporal semantics of formal logic within the VLA model.
- Trajectory-level preference optimization, specifically with a one-sided preferred-rollout anchor, effectively guides the policy to prefer satisfying executions while preserving the original natural language task capabilities.
- Explicitly encoding the formal structure of STL formulas provides a stronger conditioning signal and yields higher satisfaction rates compared to merely rendering the formulas as text within the natural language channel.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.