HERA: Bioinspired Separation of Local Fault Response and Language-Model Scheduling in Simulated Quadruped Control

Language-model inference can delay emergency supervision, but cancellation alone cannot provide an immediate physical response. HERA (Hierarchical Event-driven Reflex Architecture) separates a measurement-driven deterministic local action from language-service scheduling, with biological inspiration limited to temporal separation and local feedback. A fixed 1472-trial campaign evaluated canonical-command scheduling, simulated actuator faults and live integration. Across three workloads, the mean validated-command latency was 0.486–0.512 s with preemption, 1.476–11.101 s with single-lane FIFO and 0.543–0.575 s with two-slot serving. The original detector identified all 544 complete faults but missed all 160 half-strength faults. A separately frozen 560-trial follow-up evaluated a sustained effort-discrepancy warning without changing control. On the same 160 new half-strength trajectories, the original detector again missed every fault, whereas the warning detected all faults with mean delays of 161.0 and 155.5 ms of simulated time in clean and noisy conditions. No false onset occurred in 2400 simulated robot-seconds of new healthy exposure. This addresses conditional detection sensitivity, not improved physical recovery. In the original campaign, local support reduced the mean upright deficit but worsened some paired cases and reduced terminal speed; every physical trial met the fixed-horizon safety criterion. All 96 normal integrated supervisory commands completed 50-step holds and returned to local fallback, and local action persisted under injected transport failure. Because supervision reaffirmed the existing support array, these results establish neither additional LLM physical benefit nor general diagnosis or hardware safety.