Perceptual Compute Offloading: A Hypothesis for Sub-Millisecond Robotic Perception via Hash-Indexed Spatial Kinematic Blueprints

Vision-Language-Action (VLA) models impose inference latencies of 100–2,000ms—orders of magnitude too slow for the 50–1,000Hz control frequencies demanded by industrial manipulation. We hypothesize that Perceptual Compute Offloading (PCO), a dual-process architecture inspired by Kahneman's System 1/System 2 framework, can reduce perception-to-action latency by ≥90% for known environmental states while preserving VLA-level generalization for novel encounters. System 1 (Fast Path) isolates target objects via a zero-neural-network Kinematic ROI pipeline, computes a composite perceptual hash including a novel depth-invariant zHash, and retrieves pre-computed Spatial Kinematic Blueprints (SKBs) from a hash-indexed registry in sub-millisecond time. System 2 (Slow Path) activates only on hash miss, invoking full VLA inference and promoting new SKBs via Write-Back Augmentation. Safety is enforced through Control Barrier Functions (CBF-QP) at 1kHz with geometric gating, proprioceptive pre-tensioning, and impedance compliance mechanical fuse. The architecture achieves Six Sigma reliability through multiplicative independent verification (joint failure probability ~10⁻⁹). We present formal hypotheses with explicit falsification criteria and a standardized four-phase experimental protocol distributed as the open-source codex_lab_kit. This paper is an invitation: the first partner labs to validate or falsify these hypotheses will be co-authors on the results publication.