Reproduce “Safe Neural Control for Non-Affine Control Systems with Differentiable Control Barrier Functions” 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
Wei Xiao, Ross Allen, Daniela Rus
Researchers at MIT developed a continuous optimization learning framework that integrates differentiable Control Barrier Functions (CBFs) into neural Ordinary Differential Equations (ODEs) to ensure provable safety for non-affine control systems. This method guarantees safety for autonomous driving scenarios while operating in real-time, achieving an average minimum distance to obstacles of 1.3 meters with reduced conservativeness compared to NMPC.
Key insights
- Applying Control Barrier Function (CBF) constraints to the control derivative, rather than directly to the control input, linearizes the problem for non-affine systems, enabling efficient Quadratic Program (QP) solutions.
- Making CBF parameters differentiable and trainable allows the control policy to adapt and reduce conservativeness, leading to more optimal and less restrictive safe operation.
- Combining formal safety guarantees from differentiable CBFs with the learning capabilities of neural ODEs enables the development of controllers that are simultaneously provably safe and effective for complex non-affine dynamics.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.