Quad-VPA: Few-Step Flow-Matching Navigation over Non-Reconstructive JEPA Latents for Quadrupedal Robots in Clutter

Visuomotor navigation policies learned by imitation can match an expert offline and still fail in closed loop: in our quadrotor predecessor, Aero-VPA, behaviour cloning failed on all 60 test missions despite open-loop errors of 0.02-0.04 m/s, because the policy copied its last command. We introduce Quad-VPA, which carries the Aero-VPA architecture to quadrupedal robots and adds a terrestrial runtime shield. A single forward RGB-D frame is embedded by a Vision Transformer trained with a non-reconstructive JEPA world model; a primitive selector and a two-step conditional flow-matching actor emit chunks of planar body twists in SE(2), which a frozen, blind locomotion controller converts into joint targets, so the navigation policy never observes joint states. The deterministic shield keeps a 15 s egocentric 2.5D elevation memory with an occlusion-aware drop-off detector, enforces control-barrier conditions on a four-circle footprint through a three-variable quadratic program, turns the robot in place before it translates outside the 73° field of view, writes the intended rather than the dispatched command back into the policy's input, and stands the robot still when no admissible command exists. We prove barrier invariance per footprint circle, feasibility of standing still outside the margin, a stopping bound, and the fixed points of a copying policy under each write-back rule; specify every network, parameter and algorithm for a Unitree Go2 model in MuJoCo; and pre-register an evaluation on procedurally generated forests and the 300 BARN worlds. No experimental results are reported.