A quadcopter able to follow a person through a forest without GPS, pre-loaded maps or remote servers: that is what Purdue University's IDEAS Lab has shown in a 44-second video. Professor Aniket Bera, who leads the lab, posted it on LinkedIn last week.
In the footage the drone weaves between trees while keeping the person in frame. Every flight decision, from reading the environment to choosing the path, runs on the onboard computer. Bera argues that dropping the cloud is not a matter of convenience: a drone banking hard around a thin branch cannot wait for a distant server.
The hardest problem is the target disappearing behind a trunk. To handle it, the system includes an "occlusion prediction" step: when the person is hidden, the drone estimates where they will be when they come back into view and adjusts its path.
The subject matters beyond the lab. Under dense canopy satellite signals can be unreliable, and the US military has long sought systems that navigate without satellites. Last November Purdue announced a five-year, $1.5 million agreement with the Army Research Laboratory for drones and ground robots built for hostile terrain. Bera's post does not say whether that funding paid for the forest flights.
Bera describes the system as built entirely by his team, from the 3D-printed carbon fiber frame to the software on the chip. The graphics overlaid on the video, however, label the motion planner "SUPER" and the altitude estimate "FAST-LIO". Those are the names of two open-source projects from the University of Hong Kong's MaRS Lab. SUPER is the lab's LiDAR-guided drone, which in a Science Robotics paper early last year flew through unfamiliar terrain at more than 20 metres per second. A few weeks later the lab released the frame's design files for free.
Bera's post does not mention the Hong Kong lab, even though SUPER's code page asks anyone building on it to cite the team's papers. The IDEAS Lab's publications page has no paper on the forest demo yet.
Anyone wanting to build something similar can find SUPER's code and a full parts list, from the Livox Mid-360 LiDAR mounted on top to the Intel NUC mini PC that runs the software.
No comments yet. Be the first!