A Toronto team pairs an adaptive unscented Kalman filter with nonlinear model predictive control so a quadrotor can touch down on a moving target at a scheduled time, not an approximate one.
Drones landing on moving platforms — trucks, ship decks, autonomous-vehicle charging pads — have always been a parlor trick. They work, sort of, but the touchdown time drifts with the wind, the platform's speed, and how the drone's sensors happen to behave on a given day. A new preprint from Toronto Metropolitan University and the University of Toronto tries to tighten that variance into a schedule.
The paper, "Fixed-Time Dynamic Landing of Quadrotors using Adaptive Unscented Kalman Filtering and Nonlinear Model Predictive Control", has been accepted to the Conference on Robots and Vision (CRV 2026) in Vancouver, according to its arXiv listing. The work comes from TMU's Autonomous Vehicles Laboratory — Mohammadreza Izadi, Zeinab Shayan, and Reza Faieghi — together with Steven Waslander at the University of Toronto Institute for Aerospace Studies (UTIAS). It is a narrow engineering contribution. Its implication is wider: time-predictable landing on a moving surface is one of the open problems that has kept drone delivery and autonomous shipboard operations from feeling routine.
The pitch, in plain language, is that the drone commits to landing at a known instant — call it 2:47 pm — and the controller and the trajectory planner conspire to make that happen, regardless of how the platform is moving or how noisy the onboard sensing gets in the meantime.
Two pieces work together. The first is nonlinear model predictive control (NMPC) paired with a real-time minimum-jerk trajectory planner. Minimum-jerk references are smooth, gently-accelerating paths that are kind to motors and airframes. The planner also enforces a prescribed touchdown time during the terminal descent, so the drone is not merely aiming for a target pose but for a target pose at a specific moment.
The second piece is the Adaptive Unscented Kalman Filter (AUKF). A standard unscented Kalman filter assumes its process noise and measurement noise are roughly stationary; in the real world, neither is. The AUKF updates those noise statistics online, so the state estimate the controller feeds on does not quietly drift because the wind picked up or the vision stack started dropping frames.
The landing sequence is split into three stages — approach, pad tracking, and final descent — with the fixed-time touchdown constraint applied at the last stage. The authors frame prior NMPC landing work as leaving a gap: trajectories were not time-consistent, sensors were not noise-adaptive, and reference feasibility — whether the trajectory you ask for is one the platform can actually fly — was not analyzed.
"Fixed-time" sounds like marketing, but the authors mean something specific: a prescribed touchdown instant, not just a successful landing. The paper backs this with a reference-feasibility result: under standard tracking assumptions, the minimum-jerk references produce bounded thrust and torque commands. In other words, the scheduled landing is not a numerical trick; the motors can actually deliver it.
In simulation and hardware experiments, the authors report repeatable landings and improved platform-velocity prediction compared with EKF- and UKF-based baselines. Those numbers are author-reported, not yet independently replicated, and the paper is an arXiv preprint — the CRV 2026 camera-ready is not yet retrievable. But the trajectory-timing story is the part general-interest readers can hold: a drone that can put itself down on a moving target on schedule is qualitatively different from one that gets there, more or less, eventually.
The work is a research framework tested in lab conditions, not a deployed product. It does not say delivery fleets are ready, and it does not claim the adaptive filter solves every noise regime. Hardware-platform specifics — which quadrotor, which motion-capture or vision stack — live in the full PDF rather than the abstract, and the institutional weight is a TMU aerospace lab and one UTIAS co-author, not a consortium.
The remaining gap is the one any method-comparison preprint leaves open: turning a timing-predictable demo into a regulator-acceptable, all-weather, ship-rolling-in-a-gale primitive. The paper narrows the engineering distance to that primitive. It does not close it.