[Paper] Permissive Safety Through Trusted Inference: Verifiable Belief-Space Neural Safety Filters for Assured Interactive Robotics

Source: arXiv - 2606.02562v1

Overview

The paper introduces a new way to give autonomous robots high‑probability safety guarantees while they interact with humans. By combining belief‑space safety filters with conformal prediction, the authors show how robots can remain safe without being overly conservative, even when their internal models of human intent are uncertain.

Key Contributions

• Trusted inference framework: Extends belief‑space safety filters (BeliefSF) with a formal certification method that accounts for errors in the robot’s runtime inference module.
• Conformal‑prediction‑based safety certification: Adapts standard conformal prediction to the high‑dimensional belief space, focusing verification on regions where the inference is reliable.
• Reduced conservatism: Demonstrates that the certified filter permits more aggressive (yet still safe) actions compared to a baseline that applies conformal prediction naively.
• Empirical validation: Uses a simulated human‑vehicle interaction scenario to quantify the permissiveness gain and safety coverage of the proposed method.

Methodology

1. Belief‑Space Safety Filter (BeliefSF) – The robot maintains a probability distribution (belief) over hidden human variables (e.g., goals, willingness to cooperate). The filter checks whether a candidate control action keeps the expected future state within a safe set, given this belief.
2. Runtime Inference Module – A neural network (or other learned model) predicts the hidden human variables online, feeding the belief update.
3. Conformal Prediction Wrapper – Instead of trusting the neural predictions outright, the method builds prediction sets that contain the true hidden variable with a user‑specified confidence (e.g., 95%). This is done by:
   • Collecting a calibration dataset of inference errors.
   • Computing non‑conformity scores (how far a prediction deviates from the true hidden variable).
   • Selecting a quantile threshold that guarantees the desired coverage probability.
4. Region‑Focused Verification – The authors exploit the structure of BeliefSF: safety only needs to be certified in the part of belief space where the inference is expected to be accurate. They therefore restrict conformal prediction to that region, keeping sample complexity low.
5. Safety Certification – If the prediction set lies entirely inside the “safe belief region,” the action is approved; otherwise, the filter falls back to a more conservative safe action.

Results & Findings

Key take‑away: By narrowing the verification to the reliable inference region, the certified filter accepts significantly more actions while still respecting the same high‑probability safety bound.

Practical Implications

• Less conservative robot behavior – Autonomous delivery bots, warehouse co‑robots, or self‑driving cars can move more fluidly around people, improving throughput and user experience.
• Modular safety stack – Developers can plug the certified BeliefSF into existing control pipelines without redesigning the whole planner, preserving the “safety‑as‑a‑filter” architecture that many robotics platforms already use.
• Scalable to high‑dimensional human models – Because the verification focuses on a sub‑region, the approach remains tractable even when the belief includes many latent variables (e.g., multi‑modal intent, fatigue).
• Regulatory friendliness – The method yields a statistical safety certificate (e.g., “99 % confidence that no collision will occur”), which aligns with emerging standards for safety‑critical AI systems.

Limitations & Future Work

• Reliance on calibration data – The safety guarantee hinges on a representative calibration set; distribution shifts (e.g., new user demographics) could degrade coverage.
• Assumes bounded inference error – Extreme out‑of‑distribution scenarios may produce prediction sets that are too large, forcing the filter to become overly conservative.
• Simulation‑only validation – The experiments are confined to a simulated human‑vehicle interaction; real‑world trials are needed to assess robustness to sensor noise and latency.
• Future directions suggested by the authors include: extending the framework to multi‑robot teams, integrating adaptive calibration (online updating of non‑conformity scores), and exploring tighter probabilistic bounds beyond conformal prediction.

Authors

• Haimin Hu

Paper Information

• arXiv ID: 2606.02562v1
• Categories: cs.RO, cs.AI, cs.LG, eess.SY
• Published: June 1, 2026
• PDF: Download PDF