Safety framework for robots with partial sensors gets statistical guarantees via conformal prediction

Traditional safety guarantees in control systems often assume full state feedback, but real robots rarely have perfect information—they rely on noisy sensors, occluded views, or estimated states. In a new arXiv paper, Niloofar Jahanshahi and Mo Chen extend the control barrier-value function (CBVF) framework, which combines Hamilton-Jacobi reachability and control barrier functions, to handle partial observability. Their key innovation is applying conformal prediction, a distribution-free uncertainty quantification tool, to bound the estimation error at a user-chosen miscoverage level (e.g., 5%). This bound is incorporated into an estimator-space safety analysis, yielding a CBVF-based safety certificate specifically for partially observable nonlinear control systems.

The authors prove a finite-horizon probabilistic safety guarantee for the true system state, giving quantifiable assurances like "with 95% confidence, the system remains safe over the horizon." They also design a quadratic programming (QP)-based online safety filter for systems affine in control and disturbance, which solves in real time to enforce the safety condition against bounded disturbances. The framework is validated through a partially observable obstacle-avoidance case study, demonstrating its practical viability. This work bridges a critical gap between theoretical safety methods and real-world deployment where controllers lack complete state information—relevant for autonomous driving, drones, and assistive robotics. By relying on conformal prediction, the approach stays assumption-light and scalable compared to prior probabilistic methods.