Ivan talks about conformal reachability at CAV

​Ivan went all the way to Croatia to tell people how to put conformal prediction in a closed loop at the International Conference on Computer-Aided Verification (CAV). Doing so would let you verify autonomous systems with neural networks of any size (yes, even a VLA model like RT-2!). 

The decisive question is, to apply conformal prediction at the perception level or at the control level? 

1. If you apply conformal prediction at the perception level, you are making bounds on perception error. If so, it is wise to take into account how this error changes both over state and over time. Then state-based conformal prediction is at your service, per Ivan’s talk at the Third Workshop on Trustworthy Autonomous Cyber-Physical Systems (TACPS)​.
   
   Citation:
   • Thomas Waite, Yuang Geng, Trevor Turnquist, Ivan Ruchkin*, and Radoslav Ivanov*.
     State-Dependent Conformal Perception Bounds for Neuro-Symbolic Verification of Autonomous Systems [Arxiv] [Poster] [Slides summary] [Slides talk].
     In Proceedings of the 2nd International Conference on Neuro-symbolic Systems (NeuS), Philadelphia, PA, 2025. * Co-last authors.
     
     
2. If you apply conformal prediction at the control level, then you have a whole menu of options: single-step vs trajectory level, states vs actions. Turns out all these options lead to slightly different guarantees to bridge the gap between high dimensions (images) and low dimensions (states). These insights found their way into the poster that Ivan presented at the International Symposium on AI Verification (SAIV). 
   
   Citation:
   • Yuang Geng, Jake Brandon Baldauf, Souradeep Dutta, Chao Huang, Ivan Ruchkin.
     Bridging Dimensions: Confident Reachability for High-Dimensional Controllers [Arxiv] [Springer] [Github] [Poster 1] [Poster 2] [Slides] [Demo (w/ subs)] [Demo (w/o subs)] [Talk].
     In Proceedings of the International Symposium on Formal Methods (FM), Milan, Italy, 2024.

Photo credit: Taylor Johnson