Purdue's GPS-Free Forest Drone: What Was Shown, What Was Built "From Scratch," What Is Unverified

A short video from Purdue University is making the rounds: a small drone weaves between trees, trailing a person who jogs through dense woods. There is no GPS and no cloud connection, the post says, and everything runs on board. It is a striking clip, and it is also a clean example of why we separate what was shown from what was claimed from what has been published.

This post collects the verifiable pieces: the video from Purdue computer science professor Aniket Bera's LinkedIn, DroneXL's reporting on the open-source code visible in the telemetry, the lab's separate Army research grant, and the University of Hong Kong's SUPER repository that the telemetry points to. Where we could not confirm something, we say so.

TL;DR: claimed vs verified

What does the video actually show?

According to DroneXL, Bera posted a 44-second LinkedIn video of an autonomous drone following a person through a forest without GPS or a prior map, saying his lab "developed every layer of this autonomous drone from scratch." From the clip, DroneXL estimates the drone reaching about 7 mph, flying around 6.5 feet off the ground and trailing the person by 16 to 33 feet. Those are an outlet's readings of a video, not lab-published specifications.

Secondary posts add embellishments. Several describe 3D LiDAR mapping and obstacle avoidance "entirely on-chip," and a capability to predict where the target went when it leaves the camera's view. Those lines appear in social posts and aggregator write-ups, and the article we read that carried them offered no flight times, distances or accuracy numbers. We treat them as the lab's framing, not measured results.

How can a drone fly in a forest without GPS?

GPS signals are weak or blocked under dense canopy, and in conflict zones they are jammed. Drones in that situation navigate by LiDAR-inertial odometry: a laser scanner sweeps the surroundings many times per second while an inertial sensor tracks motion, and software such as FAST-LIO fuses both to estimate where the drone is relative to the trees it has already seen. A mapping layer such as ROG-Map turns the point cloud into an occupancy grid, marking space as free or blocked, and a planner such as SUPER picks a safe, fast path through the free space. Following a person adds a perception step that finds the target and a prediction step that guesses where they went when trees block the view. Nothing here needs GPS or a network, which is why the claim is believable in principle, even though Purdue's own performance numbers are not public.

Why does the "from scratch" line matter?

The video's telemetry overlay names two packages from the University of Hong Kong's MaRS Lab: SUPER, the motion planner, and FAST-LIO, the odometry package. A Purdue graduate student also reportedly mentioned ROG-Map, another MaRS Lab package that ships inside SUPER's repository. Bera's post does not mention HKU or the MaRS Lab, and it is unclear whether Purdue changed the code.

Here is what the public HKU repository says:

• SUPER stands for Safety-assured High-speed Navigation for micro aerial vehicles. Its demos include high-speed navigation, agile flight in dense environments and a click-and-go mode. The paper was accepted at Science Robotics in December 2024 and appeared as volume 10, number 98 (2025).
• ROG-Map is a robocentric occupancy-grid mapping framework for LiDAR-based motion planning, cited to an IROS 2024 paper, and it influenced SUPER's map representation.
• FAST-LIO is credited for localization, and the repo expects FAST-LIO2 as the odometry source.
• HKU released a hardware kit (CAD files and a parts list) in March 2025, which per DroneXL uses a Livox Mid-360 LiDAR, an Intel NUC 12 computer and an NxtPX4 flight controller.
• Per DroneXL, SUPER is under LGPL-3.0 and FAST-LIO under GPL-2.0, while the hardware kit is MIT-licensed. Check the terms before reusing any of it.

Building on strong open-source work is exactly how robotics progresses, and DroneXL's author says as much. His objection is the framing: "from scratch" and a missing credit to the lab whose planner is on screen. We agree that credit is the cheap fix. We also note that without a paper or code release from Purdue, nobody outside can tell which layers are new: the person-following behavior, the target prediction, or the integration.

What is the Army connection?

The lab is not new to this space. A December 30, 2025 article reports that Purdue's IDEAS Lab, led by Bera, received a five-year, $1.5 million cooperative agreement from the Army's Combat Capabilities Development Command Army Research Laboratory. The project, titled "Advanced Air-Ground Teaming for Autonomous Terrain Mapping and Threat-Aware Navigation," aims to build AI-enabled teams of aerial and ground robots that work together in GPS-denied, high-risk areas, with planned next steps including adaptive learning, field evaluations and larger distributed robot networks.

One aggregator describes the forest demo as part of that program. We could not find a source that says so, and the date logic is loose (the grant was announced at the end of 2025, the demo went up in late October 2026), so we leave it as context rather than attribution.

Search and rescue, or surveillance?

Both readings are on social media within hours. Some posts praise the potential for lost-hiker rescue and edge AI. Others note the obvious: a drone that follows a person under tree cover without GPS or comms is useful for pursuit and surveillance. The Army grant text makes the defense context explicit ("threat-aware navigation"), while the video itself contains a jogger and trees.

This sits in a wider conversation we have been tracking. Anthropic's Project Pilot drone benchmark asked whether language models can fly drones at all, DARPA's AI-controlled F-16 flights showed military autonomy at jet speed, and a recent report on Army AI drones overheating in soldiers' hands was a reminder that field conditions are harder than demos. A panel on autonomous warfare drew protesters in the same month.

What would make the claim checkable?

If you want to judge this kind of demo yourself, here is a short list of things to look for, which we use for any robotics video:

1. A paper or technical report with trials, failure cases and a definition of success.
2. Named hardware. Sensor, compute and flight controller, so someone can estimate whether "entirely on board" is plausible.
3. Unedited runs. Multiple runs in different terrain, including failures and cases where the target is lost.
4. Credit lines for borrowed components, which also tell you what the lab actually contributed.
5. Independent replication or at least a code release.
6. Safety and use policy. For anything that follows people, who may be tracked, and with what consent.

Today the Purdue demo meets one of these, a video, and partially the second through the telemetry.

What does this mean for builders?

Two practical takeaways. First, the barrier to GPS-free autonomous flight in clutter has dropped sharply because the building blocks are open: a LiDAR odometry stack, an occupancy-grid map, a planner and a published hardware kit are all on GitHub. If you are prototyping, start from HKU's SUPER repository and read its license terms and ROS version support (ROS1 Noetic is the top-tier platform; ROS2 is still under development) before committing.

Second, treat person-tracking as a policy decision, not just a feature. Our coverage of Anthropic's usage policy on weapons and surveillance and of Flock Safety's surveillance backlash shows how quickly tools built for one purpose get scrutinized for another. If your autonomous system touches real people, decide the rules before you ship the demo.

For readers on the AI side of robotics, Gemini Robotics 2 is a useful contrast: it focuses on language-conditioned manipulation, while this drone is classical navigation with learned elements at most. The Purdue clip does not say what, if any, learned model is in the loop.

What to watch next

• A Purdue paper, preprint or code release describing the person-following layer.
• Any acknowledgment of the MaRS Lab or an explanation of what was modified.
• Whether the lab connects the demo to its Army work.
• Independent flights with the HKU hardware kit that show how much of the behavior comes from the open stack.

We will add an update here if Purdue publishes details.

Related reading

• Anthropic Project Pilot: can AI control a drone yet?
• DARPA VENOM and AI-controlled F-16 flight
• Army AI drones overheating in the field
• Hegseth's autonomous warfare panel
• Gemini Robotics 2
• Anthropic's usage policy on weapons and surveillance

Details are accurate as of October 27, 2026, from public reporting and repositories. The lab has not published a paper, so figures may change.