Smart quadcopters find their way on their own -- without human help or GPS

and landing.

To be sure, there were sighs of despair as well. Sometimes a quadcopter would reach a point along the course and, inexplicably, hover as if dazed or confused about what to do next. After a pause, it would fly back to the starting point, having been programmed to do so if it didn’t know what to do next.

“I think it’s basically completely lost,” one researcher lamented after his team’s vehicle got close to the target in a clearing in the woods, but then took a wrong turn into another clearing and just kept going further away from the barrel. In that case, a safety pilot took over and landed the UAV so it wouldn’t be damaged, using the emergency RF link that had been installed for these experiments in the event a vehicle headed out of bounds or began flying erratically at high speed toward an object—which happened on several occasions. Undaunted by such glitches, teams would return to their tents, make some tweaks to the algorithms on laptops, upload them to the bird, and then launch again for another try.

And no, not every landing was soft. A few times the quadcopter was flying so fast, the safety pilot didn’t have time to make the split-second decision to take over. More than once that resulted in a wince-evoking “crunch”—the hallmark acoustical signature of a UAV smacking squarely into a tree or side of the hangar. Back to the team’s shade tent for some adjustments to the algorithm before uploading to a replacement craft. Each team had several UAVs on standby in their tents, and like pit crews at a raceway would quickly replace the broken bird with a fresh one to get in as many attempts as possible during their allotted 20-minute slot for each task.

DARPA notes that during each day’s morning and afternoon obstacle-course runs, at least one team was able to fly the mission autonomously, including a return to the starting point or a location close to the start—to the applause of all researchers and the test evaluators sitting under their canopies.

Success was largely a matter of superior programming. “FLA is not aimed at developing new sensor technology or to solve the autonomous navigation and obstacle avoidance challenges by adding more and more computing power,” Ledé said. “The key elements in this effort, which make it challenging, are the requirements to use inexpensive inertial measurement units and off-the-shelf quadcopters with limited weight capacity. This puts the program emphasis on creating novel algorithms that work at high speed in real time with relatively low-power, small single board computers similar to a smart phone.”

Each team brought unique technologies and approaches for outfitting their UAVs.

“I was impressed with the capabilities the teams achieved in Phase 1,” Ledé said. “We’re looking forward to Phase 2 to further refine and build on the valuable lessons we’ve learned. We’ve still got quite a bit of work to do to enable full autonomy for the wide-ranging scenarios we tested, but I think the algorithms we’re developing could soon be used to augment existing GPS-dependent UAVs for some applications. For example, existing UAVs could use GPS until the air vehicle enters a building, and then FLA algorithms would take over while indoors, while ensuring collision-free flight throughout. I think that kind of synergy between GPS-reliant systems and our new FLA capabilities could be very powerful in the relatively near future.”