Jacobian-Conditioned One-Shot Neural Reflex Control With Runtime Supervision for Emergency Evasive Maneuvers

Emergency evasive maneuvers unfold in short near-limit windows in which combined-slip saturation, load transfer, and rapidly varying friction make onset-computation budget and model mismatch critical to collision-avoidance margins. This paper proposes Neural Reflex Control (NRC), a trajectory-conditioned inverse controller that synthesizes a complete short-horizon emergency feedforward plan in a single inference at hazard onset. NRC operates inside a Jacobian-conditioned operating envelope defined on a redundancy-free two-segment parameterization (steering, net longitudinal torque, and switch time) with deterministic drive/brake decomposition. The Jacobian score is used offline for dataset curation and operating-envelope characterization; it is a local, numerical, simulator-based well-posedness indicator rather than a global guarantee. Online, lightweight runtime proxies are used to detect likely assumption violations and trigger a trusted fallback path. We refer to plan-only execution as NRC-FF and to the supervised closed-loop execution as NRC-stack. On a CarSim profile-<inline-formula> <tex-math notation="LaTeX">$\mu $ </tex-math></inline-formula> benchmark (80&#x2013;120 km/h) under the assumption-aligned protocol <inline-formula> <tex-math notation="LaTeX">${Proto-} {\mu }$ </tex-math></inline-formula>, NRC-FF achieves <inline-formula> <tex-math notation="LaTeX">$\mathrm {RMSE}_{XY}=0.059$ </tex-math></inline-formula>&#x2013;0.064 m and <inline-formula> <tex-math notation="LaTeX">$\mathrm {FSR}_{2\mathrm {s}}=86.6$ </tex-math></inline-formula>&#x2013;96.5% over a hit-radius sweep <inline-formula> <tex-math notation="LaTeX">$r_{\mathrm {hit}}\in \{0.05,0.10,0.15,0.20\}$ </tex-math></inline-formula> m. NRC-stack switches in 4.0% of rollouts while maintaining <inline-formula> <tex-math notation="LaTeX">$\mathrm {RMSE}_{XY}=0.061$ </tex-math></inline-formula> m and <inline-formula> <tex-math notation="LaTeX">$\mathrm {FSR}_{2\mathrm {s}}(0.20)=96.8\%$ </tex-math></inline-formula>. For the plan-only NRC-FF path, HIL experiments on a Jetson AGX Orin and a Hirain RT target (100 Hz CAN) show 17.16&#x2013;19.16 ms mean sample-to-latch delay (roughly two 100 Hz ticks on average), with <inline-formula> <tex-math notation="LaTeX">$\mathrm {RMSE}_{XY}=0.107$ </tex-math></inline-formula>&#x2013;0.113 m and <inline-formula> <tex-math notation="LaTeX">$\mathrm {FSR}_{2\mathrm {s}}(0.20)=94.9$ </tex-math></inline-formula>&#x2013;95.3%. Additional stress tests cover friction mismatch, out-of-family References, sensing noise/ablation, injected delay, baseline fairness, and parameter sensitivity.