Route-Guided CBF-QP Repair for
Safe Vision-Language-Action Models

Vision-language-action (VLA) models can generate task-oriented robot actions, but they do not guarantee safe execution in obstacle-rich manipulation scenes. Although CBF-QP safety filters can locally project unsafe actions into a safe set, such correction may push the robot away from the learned task trajectory and leave the VLA policy without a reliable recovery direction. We propose Route-Guided CBF-QP Repair, a post-hoc repair module that uses the short-horizon action chunk predicted by the VLA to detect blocked nominal trajectories and generate entry-exit-rejoin guidance around obstacles. The selected route is added to the CBF-QP objective as a soft guidance term, while the CBF safety constraint remains hard. On SafeLIBERO, our method improves average collision avoidance, task success, and SafeSuccess compared with AEGIS.

• VLA backbone: Both AEGIS and our method use the same OpenPI π₀.₅ checkpoint fine-tuned on LIBERO (gs://openpi-assets/checkpoints/pi05_libero). Since the VLA backbone is fixed, the comparison isolates the effect of route-guided repair.
• Baselines: AEGIS applies a post-hoc CBF-QP safety layer that projects unsafe VLA actions into the safe set. Route-Guided CBF-QP Repair (Ours) adds blocked-path detection, route candidate generation, and entry-exit-rejoin guidance to the QP objective while keeping the CBF constraint hard.
• Benchmark: Evaluations are conducted on SafeLIBERO, which extends LIBERO-style language-conditioned manipulation with obstacle constraints. For task suites, safety levels, obstacle types, and evaluation protocol, please refer to the SafeLIBERO project page.
• Evaluation protocol: Four suites — Spatial, Goal, Object, and Long — with 10 episodes per task. Metrics: Collision Avoidance Rate (CAR), Task Success Rate (TSR), SafeSuccess, and Execution Time Steps (ETS).

If you find this work useful, please cite:

@inproceedings{park2026routeguided,
  title     = {Route-Guided {CBF-QP} Repair for Safe Vision-Language-Action Models},
  author    = {Park, Chan and Hong, Sunwoo and Kim, Kirak and Park, Yoonjeong
               and Olson, Alexander W. and Pesaranghader, Ali},
  booktitle = {RSS 2026 Workshop on Trustworthy Embodied Foundation Models},
  year      = {2026}
}