The video is 44 seconds long. A small quadcopter follows a person through a forest, threading between trunks with no GPS and no prior map of the area. Aniket Bera, an associate professor of computer science who leads Purdue University's IDEAS Lab, posted it on LinkedIn on October 23 and said the team had developed every layer of the system from scratch. Two reposts on X had passed 1.7 million views combined by Monday, according to DroneXL.
The footage itself complicates that claim. A telemetry overlay runs along the video, and it labels the drone's motion planner "SUPER" and its altitude estimate "FAST-LIO." Both are open-source packages from the MaRS Lab at the University of Hong Kong, run by Professor Fu Zhang. Bera's post does not mention HKU.
Nothing here suggests the demo is fake, or that using open-source autonomy code is improper. Labs do it all the time, and that is the point of releasing the code. What the episode shows is how far apart a viral claim and the technical record can drift, and how little of the stack a lab may need to write itself to get an impressive result.
What the overlay shows
DroneXL's frame-by-frame reading of the video picks out three things. The planner field reads SUPER. The altitude field reads FAST-LIO, a LiDAR-inertial odometry package released under GPL-2.0 with about 5,300 stars on GitHub. And the flight-mode bands (Stabilized, Position, Offboard) match the standard PX4 autopilot modes. That last part is ordinary: PX4 is an open-source autopilot, and Offboard is the mode in which an external computer commands the aircraft.
DroneXL reads the speed trace as topping out near 3 meters per second, roughly 7 mph, a jogging-pace follow. HKU reported SUPER itself flying at more than 20 meters per second.
The HKU stack
SUPER, the safety-assured high-speed aerial robot, was published in Science Robotics on January 29, 2025 (vol. 10, no. 98, eado6187). The project repository lists nine authors: Yunfan Ren, Fangcheng Zhu, Guozheng Lu, Yixi Cai, Longji Yin, Fanze Kong, Jiarong Lin, Nan Chen and Fu Zhang.
HKU's description of the work focuses on a compact micro air vehicle with a 280 mm wheelbase, a thrust-to-weight ratio above 5.0, and a lightweight 3D LiDAR for long-range obstacle detection. The key idea is in the planner. On every planning cycle SUPER produces two trajectories. One stays inside space the vehicle already knows is free, so it always has a safe fallback. The other runs through both known and unknown space and is tuned for speed.
HKU reports that this approach reduced failure rates by 35.9 times compared with baseline methods while flying faster and using half the planning time, and that the vehicle flew autonomously at more than 20 m/s while avoiding thin obstacles.
The release history on GitHub is quick. The paper was accepted on December 12, 2024. A preview of the planning module, supporting both ROS1 and ROS2, went public on January 29, 2025, the day the paper came out. On March 9, 2025 the lab released the hardware components. According to DroneXL, that kit includes SolidWorks and STEP files and a parts list, including a Livox Mid-360 LiDAR, an Intel NUC 12 computer and an NxtPX4 flight controller, under the MIT license. The SUPER repository holds four main code folders: mars_uav_sim, mission_planner, rog_map and super_planner.
Taken together, the published code and hardware files give other groups a route to a close copy of HKU's research vehicle. Per DroneXL, SUPER's code is under LGPL-3.0, and the released code covers the planner, mapping and a simulator, not a person detector.
Two accounts of localization
Purdue's own descriptions of how the drone works don't fully agree. Bera's post lists visual-inertial odometry in the pipeline, meaning it estimates position mainly from cameras and an IMU. Kevin Biju Mathew, a graduate student researcher on the project, wrote in the post's comments that localization uses LiDAR-inertial odometry along with a "ROG map" occupancy grid.
Those are different sensing approaches, and the student's version lines up with the overlay. FAST-LIO is a LiDAR-inertial odometry package. ROG-Map was presented at IROS 2024 and ships inside the SUPER repository as the rog_map folder. Put the overlay next to Mathew's description and the drone looks like it is running much of HKU's published pipeline: LiDAR-inertial state estimation, HKU's occupancy mapping and HKU's planner, sitting on a PX4 flight controller in Offboard mode.
What Purdue added is not documented. The person-following behavior, the target detection, any changes to the planner, or integration work on a new airframe could all be real contributions. Bera told commenters that a prediction module estimates the person's path when they disappear behind trees. But as of Monday the IDEAS Lab's publications page listed no paper on the demo, and DroneXL found none on arXiv. Without one, outsiders can't tell which layers Purdue built and which it adopted.
What "from scratch" means in autonomy
In robotics, "from scratch" is a loose phrase. Almost no lab writes its own real-time operating system, autopilot firmware, middleware and linear algebra libraries. PX4 and ROS sit under a large share of academic drone work, and nobody expects a demo video to cite them. The usual line falls somewhere above that infrastructure: a group claims the perception, estimation, mapping and planning algorithms that define what the vehicle can do.
By that standard, the parts the overlay names are not plumbing. SUPER's two-trajectory planner is the subject of a Science Robotics paper. FAST-LIO is a LiDAR-inertial odometry package with about 5,300 GitHub stars. These are the components behind flight through a forest with no GPS and no map, and they are the parts that "developed every layer" would most naturally seem to cover.
There are benign explanations. The team may have used HKU's packages as baselines or scaffolding while building replacements, and the overlay may reflect an earlier version. The wording may have been a social-media shorthand that went further than intended. None of that has been established, and until Purdue publishes a technical description, the public record is the video and its labels.
Why It Matters
GPS-denied, map-free flight through cluttered terrain is a sought-after capability in small UAS; HKU itself points to search and rescue and disaster relief as applications for high-speed autonomous navigation. When a demo of that capability reaches more than a million viewers, the credit attached to it shapes how funders, recruiters and program managers judge which labs can deliver.
Open-source releases like SUPER exist so that other groups can build on them, and HKU also published a hardware parts list alongside the code. That openness depends on a basic norm: when you fly someone else's planner, you say so. DroneXL notes that SUPER's source files ask anyone using the code to cite the MaRS Lab's publications, and the repository's own README makes the same request. The academic incentive to credit the work is strong too, because citations are how the original authors show their release made a difference.
For readers judging autonomy claims, the lesson is practical. Look for a paper, a repository or a list of dependencies. Check what a telemetry overlay says. And treat "from scratch" as a claim that needs evidence, especially when the video on screen names someone else's software.
