XPRIZE's wildfire AI found fires in 10 minutes — couldn't stop them
The $11 million XPRIZE Wildfire handed out its awards on Wednesday, and the honest headline is the one the organizers didn't write: the detection half worked, the firefighting half didn't. Meanwhile Black Forest Labs quietly put a robot model at the top of Nvidia's leaderboard.
XPRIZE awarded $4.55 million in prizes at a ceremony in New York on Wednesday, and both $3.5 million grand prizes went unclaimed because no team could finish the job it was actually set. The competition ran two tracks. In the detection track, eight teams from four countries gathered at the New South Wales Rural Fire Service in Australia in April and were asked to spot every fire across a large swath of the country within 10 minutes; the UK-based SIRIUS Wildfire Alliance won $500,000 for a system that fuses optical and radar satellite data to model how a fire will spread, Arizona's Team Snuffed took $250,000 for detection that stays auditable and honest about its own uncertainty, and $100,000 recognition prizes went to MyRadar, DeepFire and Germany's Mayday.AI, which adapted a volcanic-ash algorithm to see wildfire smoke through cloud.
The response track was the harder test. Three finalists set up across 386 square miles of rural Alaska near Nenana in June, had to detect a high-risk ignition within 10 minutes while ignoring decoy fires, then fly out and put it out. All three detected it in time; two launched suppression drones. None extinguished the blaze. "They weren't able to put enough suppressant on it to knock it down entirely," program director Andrea Santy told Ars Technica. Anduril took first place and $1.2 million with a thermal-imaging Sentry Tower that flags ignitions, a surveillance drone swarm, and a Ghost-X drone dropping fire-retardant balls — plus a $1 million Lockheed Martin bonus for the fastest detection. Germany's Dryad Networks got $800,000 for a network of 170 solar-powered sensors that sniff the gases from smoldering ground, and AURA Foresight — a UK-Australia team that met for the first time in Alaska — took $500,000 for what its co-lead called the "Home Depot approach": cheap off-the-shelf hardware and DJI drones.
The gap is the story. A 911 call still finds most wildfires, and that averages at least 15 minutes — so an AI system that reliably flags a fire in under 10 is real progress, and the competition proved it can be done. But knowing where the fire is and having the capacity to stop it are different problems, and only the second one scales with money and aircraft rather than models. Detection is a software problem that a benchmark can settle. Suppression is a logistics problem with a physics ceiling, and $4.55 million of prizes did not move it.
Black Forest Labs released FLUX 3 Action, a 7B open-weights world action model that predicts robot movements alongside future video frames, and it immediately landed at the top of Nvidia's RoboLab-120 leaderboard. The model scored 42.9% on the benchmark — 6.1 points above Cosmos3-Nano-Policy, which needs 16B parameters to reach 36.8%, so roughly double the accuracy per parameter. Nvidia's own leaderboard carries the ranking, not just BFL's announcement. The company says the model predicts 2.13 seconds of motion at a time against 1.0 second for π0.5, and claims speedups of 1.52× to 3.95× over Cosmos 3 Nano in FP8.
Weights are live on Hugging Face in three variants — a general base model plus checkpoints fine-tuned for the SO-101 arm and the DROID dataset — under BFL's community license, which permits use with restrictions but is not an OSI-approved open-source license. That distinction matters for anyone planning to ship this commercially. FLUX 3 Action was previewed in July with the rest of the FLUX 3 family and stayed partner-only until now, which is why the leaderboard position is the news: a lab known for image models just arrived in robotics at half the size of the incumbent. The same day it shipped, we covered the other end of the robot-policy race — GPT-6 Astra drove a real car through a cone course — and the contrast is the point. Frontier labs are pushing embodied models into physical tasks from opposite directions: one with scale, one with efficiency.
What to watch: whether the unclaimed grand prizes get a second XPRIZE focused on suppression capacity, and whether BFL publishes real-robot results for FLUX 3 Action beyond its own early experiments.
Would you rather fund detection that works today or suppression that doesn't work yet — and who should pay for the difference? Tell us in the comments.
