Route-Guided CBF-QP Repair for Safe Vision-Language-Action Models
Chan Park, Sunwoo Hong, Kirak Kim, Yoonjeong Park, Alexander W. Olson, Ali Pesaranghader
Sensible, simple add-on to a recent baseline, but gains of ~2-4 points on one simulated benchmark with no variance estimates and mixed per-suite results limit impact.
Abstract
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 collision avoidance, task success, and SafeSuccess compared with AEGIS. Our project page is available at GitHub.
AI Impact Assessments
(1 model)Impact Assessment: Route-Guided CBF-QP Repair for Safe VLA Models
Core Contribution
The paper targets a specific weakness of post-hoc CBF-QP safety filters on VLA policies (here AEGIS/VLSA with π0.5). Local "closest-safe-action" projection keeps the robot safe at each step but gives no direction for going around an obstacle or returning to the task trajectory. The authors add a module that:
The idea is sensible and cleanly stated: separate hard safety from soft recovery guidance. Technically, however, it is a small extension. It combines a waypoint detour heuristic (a familiar idea from potential-field and local-planner escape strategies) with an extra penalty term in a standard QP. The framing as a "trajectory-recovery failure mode" of local CBF filters is reasonable, but local-minimum and deadlock behavior of CBF filters is well known in the control literature. The paper does not engage with that literature (e.g., CBF deadlock resolution, MPC-CBF, predictive safety filters beyond a brief citation).
Methodological Rigor
The controlled comparison is sound in principle: the same π0.5 backbone and weights are used, so the difference isolates the repair module. Several problems limit rigor:
Potential Impact
Safe deployment of VLAs is an important topic, and modular post-hoc repair that doesn't require retraining is attractive. The idea that safety filters should preserve recoverability for learned policies (out-of-distribution drift from pushing the policy off-distribution) is a useful framing. Still, the paper is a short workshop-style note: one benchmark (SafeLIBERO, simulation only), one backbone, one baseline, and modest gains. The method is likely to be cited mainly by the small group working on SafeLIBERO/AEGIS-style filters. Its hand-designed route heuristic is unlikely to become a standard component, though the "hard constraint + soft recovery term" pattern might be reused.
Timeliness
It is timely: VLA safety is an active area (SafeVLA, VLSA/AEGIS), and benchmarks for it have just appeared. It addresses a real, observable gap in the leading baseline. But the contribution is incremental on a very recent baseline, and the field is moving quickly toward learned or more principled solutions (safe RL fine-tuning, predictive CBFs, learned recovery).
Strengths
Limitations
Overall
A modest, reasonably clear engineering extension with a sensible motivation but thin evidence and limited novelty. Expected impact is low to moderate.
Generated Nov 3, 2026
Comparison History (0)
Sensible, simple add-on to a recent baseline, but gains of ~2-4 points on one simulated benchmark with no variance estimates and mixed per-suite results limit impact.