Editor's note
One pitch stayed with me this week.
Sleipner Robotics' founder said the obvious thing, that machines built unmanned from the start can be optimised for the task instead of for a human cab. He meant forest and mine robots, but the line keeps applying as I look at the rest of the issue. Pano AI's wildfire cameras are not retrofitted lookouts, they are towers built around the assumption that no one is watching live. FOR-Age does not sit in someone's hand replacing an increment borer, it runs on point clouds collected for other reasons.
Retrofit thinking gives you small efficiency gains. Unmanned-from-the-start thinking gives you whole new categories of work, things you would never have done at all if a person had to be there. I want to see more of the latter next week.
Axel
WHAT GOT ME THINKING
SingleTree's FOR-Age Reads Tree Age from LiDAR Point Clouds at ±23 Years
A new paper from the SingleTree consortium (CBE JU Horizon Europe, NIBIO-led) in Remote Sensing of Environment introduces FOR-Age, a deep-learning method that predicts individual tree age from 3D laser scanning. The team trained on roughly 1,700 point clouds from about 1,000 Norway spruce and Scots pine trees across Norway, Sweden, and Finland, from 1-year-old seedlings to 350-year-old trees. Architectures fine-tuned from ForestFormer3D's U-Net and a PointTransformerV3 trained from scratch both reached RMSE under 23 years, sensor-agnostic across terrestrial, mobile, and airborne LiDAR. Models also generalised to lower-resolution inputs when training included augmentations that simulated reduced resolution.
Axel's notes: After Deep Forestry's €3M raise two weeks back, this is another point on the trajectory I have been tracking: non-destructive single-tree measurement is going from research demonstration into something the wider stack can be built on top of. The interesting wrinkle here is the variable being estimated. Diameter, height, and species are the volume conversation, the one we have had for a decade. Age is a different conversation. Age opens up dynamics, succession, productivity by site, and biodiversity scoring in a way the volume layer does not touch. Continuous-cover management and habitat reporting both need to know how old the things in front of you actually are, and today's answer is increment borers and stand averages. If you can take that question off the field crew and run it on data you already collected, the marginal cost of asking it goes to near zero, and you start asking it at scales nobody can today. I am not putting too much weight on the headline error number yet. Two species, Nordic boreal only, and the generalisation tricks need to survive field conditions a few times before vendors should start building product on top of this. But the direction is right, and I would rather see ten papers like this than one more breathless market study.
Kodiak AI Puts Driverless Trucks on West Fraser's Alberta Log Routes
Kodiak AI (Nasdaq: KDK) will run trucks equipped with the Kodiak Driver on West Fraser Timber's log-hauling routes in Alberta later this year, moving timber from forest sites to one of West Fraser's processing facilities. It is Kodiak's first international deployment and its first push into logging, after scaling to twenty driverless trucks in West Texas's Permian Basin by the end of 2025. FPInnovations facilitated the deal. Logging roads are remote, uneven, and rough, a markedly harder operating envelope than oilfield highways, and a useful test for the claim that the Kodiak Driver is industry-agnostic.
Sleipner Robotics Lines Up Forest and Mine Pilots for Unmanned Ground Vehicles
Sundsvall startup Sleipner Robotics, founded in 2024 by ex-defense engineer Sverker Svärdby, is moving from prototype to pilot through BizMaker's Forest Business Accelerator (SCA, IBM, Sweco, RISE). Targets run from clearing and planting in forests to transport and rescue in mines. Svärdby's pitch is that machines built unmanned from the start can be optimised for the task rather than for a human cab, which keeps them simpler and cheaper than retrofitted manned platforms. Pilots are co-financed by industrial partners; an investor round is open in parallel.
Western Utilities Stack Hundreds of Pano AI Cameras Ahead of Fire Season
Arizona Public Service runs nearly 40 active Pano AI cameras (71 by summer); Xcel Energy has 126 in Colorado and aims for cameras in seven of the eight states it serves by year-end; California's ALERTCalifornia network sits at around 1,240. Pano says its tech detected 725 US wildfires last year; an APS meteorologist puts time-to-alert at roughly 45 minutes ahead of the first 911 call. Pricing is around $50,000 per camera per year. False alarms still consume staff time, and detection is where the AI stops: whether to dispatch, evacuate or hold still sits with people.