Combining small verifiers has been demonstrated to work in static environments. We extend this to dynamic environments via the OpenAI Gymnasium's Lunar Lander module.

• Proposed MAV-EI, a multi-verifier control framework that decouples fast action proposal from lightweight online verification
• Combined a Streaming Diffusion Policy with image- and state-based verifiers targeting distinct failure modes in real-time control
• Demonstrated on Gymnasium Lunar Lander that hybrid multi-verifier selection improves success rate, tightens reward variance, and reduces catastrophic failures