Our "Autonomous Mario Kart in the Wild" paper just dropped! Co-authored by researchers from Deepmind, academia & our core team, here're some lessons from approaches by research teams @SeoulNatlUni, @NUSingapore & @UTAustin A thread - 1/n 🧵

Quick recap: Earth Rover Challenge (ERC) is an open-world "in the wild" competition where research teams built autonomous navigation models to control sidewalk robots, while taking on human gamers on real-world navigation missions.

While the 1st ERC happened in Abu Dhabi at #IROS2024 , the actual missions were conducted in 8 cities around the world, from Kisumu to Singapore.

The winning AI team from @SeoulNatlUni implemented a 3-module approach, using Costmap Generation, Localization and Action Planner.

The team from @NUSingapore built a system with monocular navigation via traversability estimation with pretrained models coupled with selection of kinodynamically feasible trajectories in image space, without explicit 3D geometry reconstruction.

The team from @UTAustin built a hybrid modular approach, comprised of obstacle avoidance, terrain preference alignment, global LOC, path planning and finally Ackermann motion controller.

Congrats to all researchers & our co-authors from: George Mason U: @XuesuXiao Deepmind: @shahdhruv_ @xiao_ted @Stacormed @tingnan1986 @drzhuoxu Jie Tan FrodoBots: @micoolcho Niresh Dravin Santiago Pravisani NUS: David Hsu Meta AI: @joannetruong Georgia Tech: @naokiyokoyama0 Robotics Soc (UAE): Mohammad Alshamsi

Full paper here:

We hope to continue contributing to Earth Rover Challenge & spurring more research efforts based on such "in the wild" evaluations! More deets coming soon on what's next for Earth Rover Challenge!