A Federated Three-Layer Epistemic Architecture for Safe Robot Navigation Evidence, Planning, and Regime-Shift Recovery in a Simulated Blind-Guide Robot

We present a federated three-layer epistemic architecture for assistive robot navigation — motivated by blind-guide systems where the primary concern is protecting the user from repeated exposure to known hazards, not merely protecting the robot — and validate it through six controlled experiments in a simulated 10×10 grid-world domain. The architecture separates belief formation (Layer 1), epistemic evaluation (Layer 2), and world execution (Layer 3). All knowledge is represented as scored hypotheses over relational triplets; evidence — raw observation records with source provenance tags — is the sole unit of inter-agent transfer. This is the paper's central architectural claim: evidence, not weights or belief states, is a transferable primitive that preserves provenance, independent accountability, and deterministic recomputability in a federated epistemic system. We demonstrate five consequences of this choice under controlled conditions. (i) Epistemic gating reduces stochastic-hazard collisions by 81% on average across 120 runs (30 seeds × 4 hazard cell locations): mean 13.97 vs 2.66 collisions, SD 3.75 vs 2.93. The memoryless baseline exhibited recidivism in 100% of runs; the epistemic robot had strictly fewer collisions than the baseline in 100% of runs (V24b). (ii) Evidence federation between simultaneous robots protects an agent from a hazard first discovered by its peer, with zero personal collisions for the receiving robot (V24c). (iii) Episode-ledger transfer enables a newly deployed robot to block a hazard at its first approach — zero collisions in 84.2% of applicable runs (N=114 where Pioneer hit hazard) versus a naive peer averaging 1.51 collisions; the Heir blocked the hazard at mean step 9.4 versus the Naive robot’s first collision at mean step 45.0 — a 4.8× earlier safety response (V24d). (iv) Hypothesis versioning detects regime shifts and recovers safe navigation within 41 steps of hazard relocation via a bounded REVALIDATE_TRANSITION directive (V24e). (v) Dijkstra planning over supported transition hypotheses produces epistemically justified routes and replans correctly when a supported edge becomes contested through collision (V25, goal reached at step 112). The six primary experiments (V24–V25) are single-run over fixed seeds and reported in Section 5. Statistical confirmation analyses (30 seeds × 4 hazard cell locations for V24b and V24d; 15 seeds × 2 hazard cells for V24c) are reported inline within those sections. We make no claim about real-world robotics performance; the contribution is architectural — demonstrating that evidence federation, independent epistemic accountability, and hypothesis versioning are achievable in a clean three-layer design.

Keywords: epistemic architecture, safe robot navigation, evidence federation, knowledge inheritance, regime shift detection, epistemic planning, grid-world simulation