What Does an Agent Perceive as Possible? Perception, Passibility and the Architecture of Navigation Across Robots, Vehicles, Spacecraft and World Models

Abstract

Navigation is usually described as a stack: sense, estimate, map, plan, control. This paper asks what problem the stack solves, and whether the Architecture of Freedom Intelligence (AFI) — a theory that treats path availability as the first condition of any becoming, and intelligence as the capacity to perceive and widen that availability — adds a usable structure to it. We reconstruct how contemporary systems navigate from primary sources: classical mobile robots, Tesla's and Waymo's driving stacks, SpaceX Dragon's rendezvous guidance, video world models (V-JEPA 2, DreamerV3, Navigation World Models, the Waymo World Model) and embodied foundation models (π0.5, Gemini Robotics 1.5, GR00T N1.6, NavFoM), using 48 sources dated 2025 or later. One object recurs in every architecture: an agent-relative set of continuations treated as open. Spacecraft compute it from physics as a reachable set; Waymo materialises it as proposals checked by a validation layer; a costmap thresholds it; a world model leaves it implicit in a sampler; a vision-language-action policy emits one continuation and keeps the rest unrepresented. We then reconstruct AFI from its primary documents, including its own negative results — the failed multiplicative decomposition of its reality architecture, perception alone outperforming the perception-to-distortion ratio at graph level, the collapse of a scalar freedom field in continuous space, and the circularity of an alignment-based perception measure — and reduce each AFI concept against partially observable Markov decision processes, active inference, empowerment and maximum occupancy, reachability and viability, affordances and world models. Three moves survive reduction: perception measured as a set before choice, with real and unreal continuations separated into hidden and false paths; passibility as a layer between proposal and order; and transformation of the path space as a first-class action class. A pre-registered experiment in partially observable grid worlds tests the candidate planner against belief-rollout, empowerment-regularised and point-model planners. The pre-registered AFI planner fails by refusing to commit where the perceived set is widest; a corrected form, anchored on progress and aggregated multiplicatively across possible worlds, coincides with belief-rollout planning in every world and separates from goal-agnostic empowerment where a region rich in reachable states carries zero routes to the goal. Pre-choice perception metrics predict outcome for a reactive planner and lose predictive power for a hedging one. We conclude that AFI's contribution to navigation is a measurement protocol on perception and an architectural requirement on passibility, rather than a new predictive law, and we state the conditions under which that conclusion would change. Author ORCID: 0009-0008-6255-7724. License: Creative Commons Attribution 4.0 International.