신경신호 기반 로봇 안전 보장을 위한 실시간 이중 검증 프레임워크

Gated Uncertainty-Aware Runtime Dual Invariants for Neural Signal-Controlled Robotics Tasha Kim Oxford Robotics Institute (ORI) Department of Engineering Science University of Oxford tashakim@eng.ox.ac.uk Oiwi Parker Jones Oxford Robotics Institute (ORI) Department of Engineering Science University of Oxford oiwi.parkerjones@eng.ox.ac.uk Abstract Safety-critical assistive systems that directly decode user intent from neural signals require rigorous guarantees of reliability and trust. We present GUARDIAN (Gated Uncertainty-Aware Runtime Dual Invariants), a framework for real-time neuro- symbolic verification for neural signal-controlled robotics. GUARDIAN enforces both logical safety and physiological trust by coupling confidence-calibrated brain signal decoding with symbolic goal grounding and dual-layer runtime monitoring. On the BNCI2014 motor imagery electroencephalogram (EEG) dataset with 9 subjects and 5,184 trials, the system performs at a high safety rate of 94–97% even with lightweight decoder architectures with low test accuracies (27–46%) and high ECE confidence miscalibration (0.22–0.41). We demonstrate ≈1.7x correct interventions in simulated noise testing versus at baseline. The monitor operates at 100Hz and sub-millisecond decision latency, making it practically viable for closed-loop neural signal-based systems. Across 21 ablation results, GUARDIAN exhibits a graduated response to signal degradation, and produces auditable traces from intent, plan to action, helping to link neural evidence to verifiable robot action. 1 Introduction Neural signal-controlled robots have significant potential to improve accessibility for individuals with limited mobility but they also introduce important safety risks[32]. Maintaining runtime accuracy and implementing reliable intervention mechanisms are paramount in closed-loop systems to ensure user safety[6, 10, 17]. Recent closed-loop EEG-based assistive systems [23, 34] and AI-enabled brain- computer interfaces (BCIs) [25] show encouraging progress, but remain vulnerable to ambiguous user intent, a known challenge in shared autonomy and teleoperation[12, 19, 20]. They also suffer signal degradation during long-horizon tasks[22, 27, 29, 31] and lack formal or interpretable safety mechanisms like runtime assurance[6, 17], or shielding[7]. We propose GUARDIAN, a physiological runtime verification architecture that provides an auditable, explainable layer between neural decoding and execution. GUARDIAN acts as a safety gate that complements neurally-controlled systems, producing tracing and intervention tooling without disrupting real-time operations. Design principle and goals. GUARDIAN prioritizes the following principles: (a) conservative safety (halting over execution when neural evidence is weak or contradictory), (b) interpretability (traceable action chains from raw brain signal→intent→plan→action compatible with common symbolic planning toolchains like PDDL [13]), (c) modularity (wraps any decoder with < 1ms overhead), and (d) tunability (operators can set confidence thresholds informed by calibration analysis). Threat model and scope. Our system operates with non-invasive neural signals in assistive robot domains, oftentimes challenging due to noise interference, non-stationarity–within-session accuracy drop and between-session variance, neural artifacts, and subject fatigue[9, 29]. We address confidence 39th Conference on Neural Information Processing Systems (NeurIPS 2025) Workshop: Embodied and Safe- Assured Robotic Systems. arXiv:2511.20570v1 [cs.RO] 25 Nov 2025 miscalibration, where decoders show high confidence despite poor accuracy[15, 26]. When there are no violations, our monitor gates actuation and operates under standard action safety protocols[2, 3, 5]. In assistive robotics, hardware can pose physical risks for users when there is uncontrolled device activation or delayed shutdown[5, 6]. Attacks or adversary threats are outside our scope. Contributions. We make the following contributions: (I) a safety-centric framework achieving high safety rates under severe signal degradation, (II) a verifiable neuro-symbolic pipeline that formally links EEG distributions to symbolic goals, and (III) a practical lightweight architecture with minute computational overhead enabling deployable and trustworthy neural signal-based control. 2 GUARDIAN: Formalization and Algorithm Figure 1: System architecture overview. Let xt ∈RC×W denote the EEG window at time t (C channels, W samples). A decoder fθ outputs a categorical poste- rior pt ∈∆3 over the action set H = {GRASP, RELEASE, MOVE_TO, ROTATE} corresponding to left hand, right hand, feet, and tongue motor imagery (MI), respectively. These actions enable natural control (positioning, grasping, releasing, orienting) for EEG-based manipulation. We treat pt as a 4-dimensional probability vector over H, i.e., pt ∈∆3 ⊂R4 (3-simplex). Let u denote the uniform distribution on H, i.e., u(h) = 1 |H| = 1 4 for all

…(Full text truncated)…

This content is AI-processed based on ArXiv data.