AI Racing FPV Drone Full Send! A.I vs Humans - University of Zurich

This is an artificial intelligence-piloted drone flying insanely fast around a race track.

Video by BMS.

This is a very quick preview of the University of Zurich AI drone using the Vicon camera tracking system. Sharing this because it was so awesome to see a JS-1 Drone flown to its full potential! Absolutely Insane.

Will be following this video up with way more details once we get back home to Australia and will also cover their Vision-based race drone along with the reasoning behind this sort of research.

Hereโ€™s some information posted by Tattu Lipos, who sponsored the event,

๐—ง๐—ต๐—ฒ ๐—›๐˜‚๐—บ๐—ฎ๐—ป ๐˜ƒ๐˜€ ๐—”๐—œ ๐——๐—ฟ๐—ผ๐—ป๐—ฒ ๐—ฅ๐—ฎ๐—ฐ๐—ฒ, organised by the University of Zurich Robotics and Perception Group, took place this past weekend at Swiss Drone Days, and as one of the partners, Tattu was delighted to be there to witness this historic moment!

UAV is currently being used extensively in several transportation fields, but planning time-optimal trajectories at the actuation limit through multiple waypoints remains an open problem. The Lab, therefore, turned its attention to FPV in an attempt to load powerful algorithmic systems onto this high-speed sport to achieve the fastest possible time allocation.

A week before the race the lab invited three world champions, Thomas Bitmatta, Alex Vanover and Marvin Schรคpper to start practicing, collecting eye movement data and other optimizations of the AI algorithm system. Here, the track and the drones were built for AI, complete with 36 cameras and sensors, the equivalent of adding 36 powerful brains to the meticulous algorithms!

During this time, the pilots have probably race directly against the AI 2-3 times. Marvin Schรคpper said: โ€œI think on one lap human can be faster. But on 3 or more laps the AI will win. times behind each other because they can just repeat one lap.โ€

Also in practice, our human pilots are constantly adapting their strategies and upgrading their equipment to get faster times. Thomas Bitmattaโ€™s father Pual told us : โ€œOn the track we were doing, Thomas got down to a 6.26 on the Tattu 4.0, however on the 5.0 he managed a 6.066!!!โ€

The challenge started last Saturday with a 1V1 challenge in the morning between the AI and the pilot for the fastest lap time. After several rounds of attempts, Marvin Schรคpperโ€™s fastest lap was 5.6S, Alex Vanover 6.7S and Thomas Bitmatta 6.3S. The AI finished in 5.3S the first time, 5.1S the second time and crashed the third time. Despite the seemingly disparate best times, Alex Vanover came up with a brilliant way of blocking his quad in the middle of the traversing gate, where the AIโ€™s sensors detected a fault and could not get through then crash!! Victory!

The second round in the afternoon was mainly an image recognition session. This time the AI changed the quad from being covered with countless sensing cameras to having 1 camera on the head, which is equivalent to going from 36 brains to 1 brain. Alex Vanover and Thomas Bitmatta both achieved 5.6s and Marvin 5.44s. The AI slowed down due to the light recognition, with a best lap time of 5.6s. AI tried to set a new record again, but unfortunately hit a gate and the battery cut and started to catch fire and smoke. The organisers immediately evacuated the crowd.

The AI machine was taken to the lab for repair and no solution could be found at the moment. There were five AI drones in total, all of which had previously suffered damage of varying degrees. The professors had an emergency internal meeting and decided to have another battle between the three pilots in order to give the audience a good time. Alex Vanover won several times and he excitedly ran to the audience waving and cheering.

Shortly afterwards, the assistants checked the induction system again, the track was rebuilt and it looked as if the professors had come up with a solution. After a lot of hard work by the lab members, the AI system was finally fixed! Round 4 begins! Alex Vanover has a flight time of 5.288s, Thomas Bitmatta 5.48s and the AI has a time of 3.53s after a successful take-off!!!

A congratulatory result! Congratulations to the whole AI team! And a big round of applause to the human pilots, Tattu is proud of each of you! This is not only an advancement in technological algorithms but also humans pushing their limits with technology! It was also a duel between the present and the future, and whoever won is proud of it.

Human-controlled racing drone compared to A.I controlled drone - BMS Racing.

BMS JS-1 (Human) vs VICON JS-1 (AI) drones side by side. Note the AI doesnโ€™t need a camera because it will have 35 external cameras. Either way, a JS-1 will win this battle :heart_eyes:
BMSweb on Instagram: "BMS JS-1 (Human) vs VICON JS-1 (AI) drones side by side. Note the AI doesnโ€™t need a camera because it will have 35 external cameras. Either way a JS-1 will win this battle ๐Ÿ˜ #bmsjs1 #bmsracingJS1 #squishisbetterthenstretch ๐Ÿ˜œ #bmsweb #racing #drone #droneracing #tbs #teamblacksheep #impulserc @tmotorfpv @hqpropzhong @tattulipos @impulserc_fpv @bmsthomas_ @missionfoodsanz #missionfoods @bmsthomas_ @radiomasterrc @swissdronedays"

Time-Optimal Online Replanning for Agile Quadrotor Flight

A video shared by the University of Zurich explaining how they developed the A.I. Racing system.

How long do you think until A. I drone technology can outpace human pilots without the need for external cameras/processors, with just an onboard camera?

1 Like

I remember just a few years back when an A.I drone could barely make it through a gate. The technology has come a long way in a short time. Crazy to think what it might be capable of in another few years.

1 Like

Someone call MCK

1 Like