ORBIT: Guided Agentic Orchestration for Autonomous C-to-Rust Transpilation
New framework achieves 100% compilation and 91.7% test success on large codebases, nearly eliminating unsafe Rust.
Researchers from Stony Brook University and AWS have developed ORBIT, a novel framework that uses guided AI agents to automate the complex task of translating legacy C code into safer Rust. Unlike previous LLM-based approaches that struggle with large codebases and complex dependencies, ORBIT employs a multi-agent system that dynamically navigates a project, constructs a dependency-aware translation graph, and coordinates specialized agents for interface generation, function mapping, and iterative verification. This orchestrated approach tackles the core challenges of hallucination and context limitation that plague simpler translation tools.
ORBIT was rigorously tested on 24 real-world programs from the CRUST-Bench, with over 91% exceeding 1,000 lines of code. The results are striking: it achieved 100% compilation success and 91.7% test success in both expert-guided and fully autonomous modes. This performance substantially outperforms existing tools like C2Rust and the baseline CRUST-Bench system. Crucially, ORBIT's translations drastically reduce the number of `unsafe` code blocks—a key Rust feature for memory safety—bringing them close to zero. The framework also demonstrated competitive results on the challenging DARPA TRACTOR benchmark, proving its scalability for large, critical systems.
- Achieved 100% compilation & 91.7% test success on 24 large CRUST-Bench programs
- Uses multi-agent orchestration to dynamically navigate dependencies, beating static-context LLM approaches
- Reduces unsafe Rust code blocks to nearly zero, a major win for memory safety
Why It Matters
Automates the risky, expensive process of modernizing critical legacy infrastructure, making systems fundamentally safer.