Multi-step reachability with asymmetric conformal prediction for safety-critical autonomous systems
Loading...
Authors
Chaabra, Shriansh
Issue Date
2025-12
Type
Electronic thesis
Thesis
Thesis
Language
en_US
Keywords
Electrical engineering
Alternative Title
Abstract
Data-driven surrogate models have been used to approximate nonlinear system dynamics that are otherwise unknown. To provide safety guarantees for a system with such unknown dynamics, prior work has proposed a one-shot formulation that computes conformal flowpipes for entire multi-step trajectories in a single forward rollout, offering distribution-free coverage for surrogate-based reachability.This thesis advances the idea to scenarios where the system’s state must be re-evaluated as it evolves over time. Rather than predicting an entire trajectory in one pass, the method builds multi-step reachability tubes by repeatedly rolling the surrogate model forward from the current state across a limited prediction horizon. At each step, the forecast is recalibrated using conformal residuals collected at different horizons, producing adaptive scaling factors that reflect how the model’s uncertainty grows with each successive propagation.
In addition, symmetric uncertainty padding is replaced by asymmetric conformal quantiles, learned separately for the lower and upper residual tails, so that safety margins expand only in directions where prediction error is significant.
The resulting method produces adaptive probabilistic flowpipes that tighten or widen in response to the observed dynamics, providing controllers with earlier warning horizons and more compact, direction-aware safety envelopes for real-time decision making.
Description
December2025
School of Engineering
School of Engineering
Full Citation
Publisher
Rensselaer Polytechnic Institute, Troy, NY