Safe online adaptive refinement

Beyond safe execution.

SOAR-Flow helps an offline-initialized driving policy improve online while every reference is checked before execution. Watch the complete simulation and UGV evidence below.

Zeqiao Li · Yijing Wang · Haoyu Wang · Zheng Li · Zhiqiang Zuo

▶ Watch the 2:54 demo arXiv link coming soon Reference code Open the repository README
SOAR-Flow · final 1080p demonstration with English narration and subtitles.

Follow the evidence

Framework: Flow reference, CBF–MPC certification, calibrated values, and adaptive guidance.
Simulation: policy-only failure, protected execution, online improvement, and near correction-free driving.
UGV: blocked-side bypass and the successive-turn 8-shaped maneuver.
Held-out cornering transfer, comparison, and final execution.

What SOAR-Flow contributes

Protected interaction

CBF–MPC certifies policy references before they reach the controller and records the certified transition for learning.

Value calibration

Gaussian return variability smooths the Cal-QL calibration pressure and reduces policy degradation during online updates.

Adaptive guidance

Corrections teach feasible recovery while intervention-to-go prediction gradually reduces auxiliary guidance as dependence falls.

Reported results

352.3 ± 2.1
final simulation return
0.72 ± 0.12%
simulation correction rate
95%
held-out cornering safe success