Glonce

Runtime fix restores safety in 73% of ACC perception failures

researchThe Glonce editorsarXiv🔥72Glonce rating72Impact70Novelty75Relevance95Surprise50Players20Credibility75

Machine-learning perception is now central to many autonomous systems, but unlike traditional sensors, the boundary where it does or does not work reliably is poorly characterized. When ML perception gets it wrong, the downstream control system can act unsafely or, in the opposite failure, become overly conservative.

Researchers propose a two-step strategy for correcting ML-based state estimation without retraining the perception model itself. First, an offline computation characterizes the uncertainty in the perception module's state estimates, using what the authors call preimages of perception contracts: regions that capture where and how the estimate can be trusted. Second, at runtime, a risk heuristic uses those characterized uncertainties to pick a particular state from the set of possibilities and feed it to the control decision, rather than blindly trusting the raw perception output.

The method was evaluated through extensive simulation on vision-based adaptive cruise control (ACC) systems, across different weather conditions and road scenarios, using the Yolo object detector and the LaneNet lane detector for perception. The test set was built from 45 ACC scenarios in which the original, uncorrected perception-based control system using Yolo and LaneNet had led to safety violations. Applying the runtime correction preserved safety in 73% of those 45 scenarios. The remaining 27% could not be recovered, because in those cases the construction of the preimages of perception contracts was not fully conformant. The authors also report that the correction is not overly conservative: on average, corrected scenarios saw only a 2.8% increase in completion time compared with the uncorrected baseline, with the interventions described as mild.

Key facts

  • The method combines an offline step, characterizing perception uncertainty via preimages of perception contracts, with a runtime risk heuristic that selects among the uncertain states to drive control decisions.
  • It was tested on 45 ACC scenarios where the original Yolo and LaneNet based perception-control system had already produced safety violations.
  • The runtime correction preserved safety in 73% of those 45 scenarios.
  • The remaining 27% could not be recovered because the preimages of perception contracts were not fully conformant in those cases.
  • Corrected scenarios experienced only a 2.8% average increase in completion time versus the uncorrected baseline, with mild interventions.

Why it matters

Traditional sensors have well-understood failure modes; learning-based perception does not, and the authors state plainly that the boundary of when ML perception works is poorly characterized. That gap matters because incorrect perception can push an autonomous system into unsafe action or, just as problematic, into being needlessly cautious. This work targets that gap directly, at runtime, without requiring the perception model to be retrained or replaced.

Who it affects

The immediate audience is engineers building safety-critical autonomous systems that rely on ML-based perception, specifically vision-based adaptive cruise control using object detectors like Yolo and lane detectors like LaneNet. The approach is framed as a general strategy for correcting ML-based state estimation, but the reported evaluation covers only ACC modules, not other autonomous-system categories.

How to use it

The strategy runs in two stages. Offline, a computation builds preimages of perception contracts that characterize the uncertainty coming out of the ML perception module. At runtime, a risk heuristic uses that characterization to choose a specific state from among the uncertain estimates, and that chosen state, rather than the raw perception output, drives the control decision. It sits as a correction layer on top of an existing perception-and-control pipeline. The source gives no pricing, licensing or availability details.

How solid is it

The evaluation is extensive simulation across multiple vision-based ACC modules, different weather conditions and road scenarios, though the source does not enumerate which specific conditions and scenarios were used. The test set itself is a hard one by construction: all 45 scenarios were cases where the uncorrected system had already caused a safety violation. The correction preserved safety in 73% of them. No real-world or on-road testing is reported, and no absolute completion-time baseline is given, only the relative 2.8% increase.

Risks and caveats

The method does not close the gap entirely: 27% of the 45 violating scenarios remained unrecovered, and the authors attribute this specifically to cases where the preimages of perception contracts were not fully conformant, an acknowledged limitation of the construction itself rather than of the test conditions. The source also does not describe in detail what the runtime risk heuristic actually computes beyond its name, and results come from simulation only, with no author names, affiliations or publication date given in the text.