Racing Without a Map: Bearings-Only Guidance and Optical Looming for Autonomous Drone Racing

Research question

How should an autonomous racing drone traverse a single fixed but unsurveyed course when pose, odometry, official gate state, and surveyed gate coordinates are all blocked? The admissible agent receives a monocular 640 × 360 camera, inertial measurements, timing signals, and its own command history.

Literature findings

The report surveys roughly 400 works spanning 1958–2026 and applies explicit admissibility filters. Only four published systems operate with camera and IMU but without odometry or a map; the leading monocular racing systems still depend on a track map and cannot transfer unchanged.

Four mature research lines already solve parts of the problem—optical time-to-contact guidance, bearings-only gap traversal, image-based visual servoing, and image-space flight through openings—but the survey finds no prior system joining them for autonomous racing.

A control-rate limit before policy design

Using a conservative 0.109-second surrogate control period, a drone moving at 8 m/s would travel 0.87 meters between decisions. Compared with a 0.61-meter gate half-width, that surrogate yields a 5.6 m/s clearance bound and shows why control rate must be measured before policy design. The report notes that a roughly 30 Hz live loop would raise the bound to about 18 m/s and make it nonbinding.

Proposed mapless system

The proposed architecture combines a deterministic bearings-only and optical-looming expert, a drag-aware attitude estimator, gate-corner reprojection in an error-state filter, and a speed-conditioned recurrent policy. Multi-fidelity Bayesian optimization specializes the system to the fixed course under a three-attempt qualification budget.