An arXiv preprint reports a camera-only neural network for predicting continuous yaw corrections—the direction and amount of a horizontal turn—from single views extracted from 360-degree video. Separately, its offboard semi-autonomous test recorded a successful traversal of a previously unseen indoor laboratory, including passage through a doorway without collision.
The research question is framed around recovery guidance for micro-UAVs in situations where GPS is unavailable. The practical test, however, is described as an offboard setup using a DJI Tello EDU, an external computer and Wi-Fi, so the reported result should be read as a laboratory validation of that setup rather than as an onboard test of the custom recording drone.
The document is an arXiv preprint labeled v1 and dated 20 August 2026.
From a spherical recording to a turn target
The source footage came from a custom micro-drone equipped with a 360-degree camera. The pipeline converted equirectangular frames—the flat format used to store a spherical view—into planar front views and dynamically generated image-label pairs for the yaw task.
For each equirectangular source frame, the generator created 10 planar views from the front yaw range. Drone recordings also supplied reused views from the back range after those views were mapped to the front range.
The processed set covered 10 scenarios and 156 raw videos: 46 from the drone and 110 recorded manually. It contained 11,058 equirectangular frames, including 5,806 indoor and 5,252 outdoor frames; 7,051 frames came from drone recordings and 4,007 from manual recordings.
Of those frames, 9,953 were used in training and validation, while 1,105 drone-only frames formed the held-out test split. Dynamic view generation expanded the material to 181,090 planar samples, including 40,070 derived from manual recordings and 141,020 from drone footage.
A contest among nine designs
The task was supervised regression: the network was trained to estimate a continuous yaw correction from each planar RGB view. Training used mean squared error, which penalizes larger misses more heavily, stochastic gradient descent, 100 epochs and batches of three, with a 90% training and 10% validation split.
The study compared nine CNN variants across five design families: tanh, ReLU, dropout, batch normalization and leaky-ReLU.
The strongest score came from leaky-ReLU
The strongest held-out regression result came from leaky-ReLU with a negative slope of 0.05. It reached an R2 score of 0.8074; on the model’s internal output range, MSE was 0.0649 and RMSE was 0.2548.
On the degree scale, the same model had an RMSE of 15.28 degrees and a mean absolute error, or MAE, of 10.146 degrees. RMSE gives more weight to larger errors, while MAE summarizes the average absolute size of the misses.
In the tested architecture comparison, ReLU had better validation loss and test metrics than tanh, dropout variants performed worse, and leaky-ReLU had the best overall test performance among the designs examined.
A separate three-class classification check used left, straight and right labels as a sanity check. Its best run reached 0.9201 accuracy, while the mean over 10 runs was 0.9099.
A practical check beyond the dataset
The speed tests measured the forward pass, or the calculation of one prediction. On a high-end workstation GPU, reported times ranged from 0.166 to 0.498 milliseconds, with throughput of 2008 to 6024 predictions per second. On an embedded Jetson Orin GPU, times ranged from 1.486 to 1.528 milliseconds and throughput from 654 to 673 predictions per second.
On a low-power Raspberry Pi 4-class CPU, the forward pass took 45 to 112 milliseconds, with throughput of 9 to 22 predictions per second.
The practical demonstration used a DJI Tello EDU rather than the custom 360-degree recording platform. Inference ran offboard through an external computer and Wi-Fi, and the previously unseen indoor laboratory was traversed successfully, including passage through a doorway without collision.
The hard scenes
The largest reported errors involved high-dynamic-range lighting and specular reflections, along with texture-poor scenes where multiple escape directions looked visually plausible.
That failure pattern is especially relevant to a single-view continuous prediction task: the difficult scenes could present more than one visually plausible escape direction, while lighting extremes and reflections were also reported as problem conditions.
The evidence is therefore bounded by the reported held-out split, timing benchmarks and offboard laboratory setup. The research question concerns GPS-denied recovery guidance, while the practical result came from the DJI Tello EDU with an external computer and Wi-Fi, not onboard inference on the custom recording platform.
A preprint with public materials
The authors state that both the dataset and the training source code are publicly available.
Paper data and sources
Original title: A 360-Degree Vision Dataset for Learning Yaw Control on GPS-Denied Micro-UAVs in Disaster-Response-Relevant Environments
Authors: Niklas Voigt, Hartmut Surmann
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text