WEBVTT

00:00:00.000 --> 00:00:05.200
This work is called SOAR-Flow — Safe Online Adaptive Refinement.

00:00:05.200 --> 00:00:10.800
Safe online fine-tuning: learn without danger, and need fewer safety corrections.

00:00:10.800 --> 00:00:13.600
The Flow policy proposes a driving reference.

00:00:13.600 --> 00:00:16.000
CBF–MPC certifies it and corrects risky references before execution.

00:00:16.000 --> 00:00:17.800
Protected returns are used for value calibration.

00:00:17.800 --> 00:00:25.000
Paired corrections and future-intervention prediction guide the policy and gradually reduce auxiliary guidance.

00:00:25.000 --> 00:00:29.000
Simulation uses SafeMetaDrive with new road layouts and obstacle configurations.

00:00:29.000 --> 00:00:33.000
The initial policy-only execution fails in this new layout.

00:00:33.000 --> 00:00:41.000
With CBF–MPC enabled, the same policy is protected and passes safely.

00:00:41.000 --> 00:00:49.000
After online updates, the intermediate policy needs only occasional correction.

00:00:49.000 --> 00:00:54.000
At one million steps, the key segment is almost correction-free.

00:00:54.000 --> 00:01:02.000
With high-level correction disabled, the refined policy still completes the task.

00:01:02.000 --> 00:01:12.000
Protected experience is converted into policy capability.

00:01:12.000 --> 00:01:22.000
Return recovers while protected training cost and correction rate decrease.

00:01:22.000 --> 00:01:29.000
We next use a real uncrewed ground vehicle and collect left- and right-passing demonstrations.

00:01:29.000 --> 00:01:39.000
We move the front obstacle to the right and block one demonstrated route.

00:01:39.000 --> 00:01:50.000
Early runs turn too late, so CBF–MPC corrects the reference. Later runs prepare earlier and need fewer corrections.

00:01:50.000 --> 00:02:01.000
The eight-shaped task tests two successive turns in opposite directions.

00:02:01.000 --> 00:02:10.000
Early runs require correction; the intermediate behavior approaches the safe bypass pattern.

00:02:10.000 --> 00:02:19.000
The late policy completes the maneuver without high-level CBF correction, while low-level MPC tracking remains active.

00:02:19.000 --> 00:02:32.000
On the held-out cornering task, the frozen policy clears both obstacles without further updates.

00:02:32.000 --> 00:02:44.000
Compared with fixed CBF–MPC, it achieves higher success, shorter completion time, and larger minimum clearance.

00:02:44.000 --> 00:02:54.000
SOAR-Flow combines safe interaction, calibrated values, and fewer corrections — a policy that keeps improving inside the safety boundary.