Reproduce “Safe LLM-Controlled Robots with Formal Guarantees via Reachability Analysis” 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
Ahmad Hafez, Alireza Naderi Akhormeh, Amr Hegazy, Amr Alanwar
A safety assurance framework allows large language model (LLM)-controlled robots to operate with formal safety guarantees by employing data-driven reachability analysis, which avoids the need for explicit analytical models. The framework successfully prevented all collisions in simulated TurtleBot3 and real-world JetRacer experiments, enabling robust obstacle avoidance.
Key insights
- Formal safety guarantees can be provided for LLM-controlled robots, even with their probabilistic outputs and nonlinear dynamics, by overapproximating reachable sets using a data-driven approach.
- Efficient set representations (zonotopes and constrained zonotopes) enable real-time computation of reachable sets and differentiable collision checking, allowing for on-the-fly adjustment of LLM-generated control actions.
- Integrating a zero-shot LLM with a formally verified, data-driven safety layer provides a practical way to combine high-level natural language understanding with low-level safety assurance in autonomous systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.