Most planners think in world coordinates; Skydio's claim thinks in pixels. Its grant US11592845B2 ("Image space motion planning of an autonomous vehicle," issued February 28, 2023; inventors Ryan David Kennedy, Peter Benjamin Henry, Hayk Martirosyan, Jack Louis Zhu, Abraham Galton Bachrach, Adam Parker Bry) plans an autonomous drone's motion directly in image space — reasoning about where to go in the camera frame rather than first reconstructing a full metric 3D world and planning there.

Claim 1 makes "image space" a precise, recited construct rather than a slogan. The system processes a captured image "to generate a cost function map corresponding to an image space of the image, the cost function map associating a cost value with each of a plurality of regions of the image," where each cost value "is indicative of a level of risk associated with navigating in an area of the physical environment corresponding to the region of the image." So the cost field lives on the 2D image grid: every pixel-region carries a navigation-risk cost. The planner then "identifies…a particular region of the image that is associated with a higher cost value" and — this is the inventive verb — optimizes "a planned 3D trajectory…by adjusting the planned 3D trajectory such that a projection of the planned 3D trajectory in the image space of the image avoids the particular region." The 3D path is bent by projecting it into the camera frame and steering its projection away from high-cost pixels. Planning stays anchored to the sensor's native 2D representation; the world model is never fully reconstructed.

How the per-region cost is computed is the part that ties this to a real flying camera. Claim 7 (and its method twin, claim 15) defines the cost: "estimate depth to a physical surface in an area of the physical environment corresponding to the particular region…and determine a level of confidence in the estimated depth." Cost rises both where surfaces are close and where the depth estimate is uncertain. That second factor is the clever, safety-aware bit: the drone is pushed away not only from things it sees clearly as near, but from regions where it cannot be confident what is there. Low-confidence depth is treated as risk — exactly the right bias for a monocular-or-stereo flying camera whose depth estimates degrade on texture-poor or distant surfaces. Claim 5/13 lets that risk be assigned by "a machine learning model," so the cost map can be learned rather than hand-engineered.

“An autonomous vehicle that is equipped with image capture devices can use information gathered from the image capture devices to plan a future three-dimensional (3D) trajectory through a physical environment. To this end, a technique is described for image-space based motion planning.”— U.S. Patent No. 11,592,845 source

The claim also handles the fact that a drone is moving while it plans. Claims 8 and 9 add receiving "an additional image…while the vehicle is in motion," processing it to "update the cost function map," and re-optimizing the trajectory against the updated map — a closed perception-planning loop that refreshes the image-space cost field as new frames arrive. That is what makes the approach viable for real-time obstacle dodging: the planner is not solving once against a static reconstruction, it is continuously re-projecting and re-steering against the latest frame. Claim 2 closes the actuation loop, "process the optimized planned 3D trajectory to generate control commands" and "control the propulsion system…to maneuver along" it, and claim 4 fixes the embodiment as "an unmanned aerial vehicle (UAV)."

Keeping planning close to the sensor is the design philosophy worth filing. G06T 7/246 (visual tracking) and 7/277 (filtering of tracked features) mark the image-domain reasoning; B64C 39/024 marks the drone airframe and G05D 1/106 the 3D flight control. Image-space planning is computationally lighter and degrades more gracefully than full 3D reconstruction — a sound choice for a power- and compute-constrained flying camera that has to dodge obstacles in real time. It is a different camp from the map-heavy ground-AV approach, and fencing it stakes out that lighter-weight philosophy.

From a portfolio angle, this is a deliberate multi-grant family: the identical title and inventor set recur across US11592844B2 (2023) and US11787543B2 (2023), a tight cluster of continuations re-fencing image-space planning. Skydio — the leading US autonomous-drone maker — built this family as the moat under its obstacle-avoidance autonomy, the capability that differentiates its product. The three independent forms here (system claim 1, method claim 10, vehicle claim 18) already spread the same image-space-cost-and-projection method across device and method.

Caveats. Image-space and visual-servoing-style planning has robotics prior art; the grant turns on the specific method in claim 1 — the per-region image-space cost map, the depth-plus-confidence cost, and the trajectory-projection avoidance step — not on the concept. With a three-grant family, the enforceable scope is the union of the live claims; read the latest. An accused planner likely has to reason about cost in the image frame and steer a projected trajectory to read on this — a ground vehicle that reconstructs a full metric occupancy grid and plans in world coordinates would sit outside the claim, which is exactly the line the drafting draws.

For the file: a vision-native drone-planning family central to Skydio's autonomy, distinguished by an image-space risk-cost map (weighted by depth confidence) and trajectory-projection avoidance. Pull US11592845B2 and its 2023 siblings in the patent record, read the newest claim 1 with claim 7 for the cost definition, and file it as the drone camp's answer to ground-AV's map-heavy planning.