CUPI at Cornell University
Passing Every Gate Without Learning: A Deterministic Policy for AI Grand Prix VQ1
How CUPI cleared all six AI Grand Prix VQ1 gates with deterministic dead reckoning, monocular corrections, and no learned policy component.
- Cleared all six Virtual Qualifier 1 gates in order
- Used deterministic control with no learned component
- Combined monocular gate-size and bearing corrections with an admissible scored-event timing signal
CUPI Technical Report · 2026 · Updated
Read the full 3-page technical report (PDF)
Result
The policy passed Virtual Qualifier 1 of the Anduril AI Grand Prix, clearing all six gates in order. It is deterministic: from the same start state and configuration it produces the same command sequence, making scarce qualification attempts reproducible.
Policy architecture
The controller integrates reported body-frame velocity and attitude into a position estimate, aims through the released six-gate course geometry, and publishes collective-thrust-and-body-rate commands at 50 Hz. Three independent correction channels keep accumulated drift and gate bookkeeping bounded.
- Apparent gate size provides an absolute position correction against known gate geometry.
- Gate bearing adds a direct visual-servo correction during the final approach.
- The elapsed-time-since-last-gate signal resynchronizes the active gate index after a scored pass.
Deploy boundary
The deployed policy reads only the admissible observation frame and camera image. It does not consume privileged simulator state, privileged scorer outputs, collision flags, or oracle pose; it does use the admissible elapsed-time-since-last-gate signal. A content-addressed manifest fixes the course revision, controller, and tuned parameters for every promoted run.
Why VQ2 requires a different system
Virtual Qualifier 2 removes attitude, local position, odometry, and gate telemetry. Those restrictions invalidate the mapped dead-reckoning approach rather than merely requiring new tuning. CUPI’s successor work therefore guides from image bearings and optical looming without a surveyed map.