Applied Intuition

Making a machine intelligent is hard. Making one system that works across cars, drones, trucks, and industrial equipment is the much harder problem we’re solving. https://lnkd.in/gbSXQ_NK

How does labeled data become a model back on the road? Shiyan Xu, an engineer on our Data Engine team, breaks down how we take labeled data through curation, scenarios, data loading, training, and evaluation, all the way back to the fleet, across hundreds of petabytes of multimodal data from automotive, trucking, mining, and robotics fleets. Closing that loop comes down to a few key design choices: 👉 Choosing segments to feed the model: Curation runs selection criteria, like nighttime driving with a cut-in under two seconds, against a curation table hundreds of columns wide. Scenarios no metric or label covers can be found by searching clip embeddings with text or an example clip. 👉 Building scenarios for training: Each scenario bundles aligned frames, labels, map, and route into one row. Lance’s blob v2 API makes each scenario’s blob payload easy to work with: written and versioned with the row, read only when needed. Data evolution adds new fields without rewriting existing files. 👉 Turning scenarios into model inputs: Lance’s random access lets training sample rows non-sequentially in small, scattered batches, with each batch fetched as a single take. Checkpointing lets a restarted job pick up where the sampler left off instead of replaying. 👉 Training in two stages: Imitation learning first teaches the model to drive like the fleet’s drivers, then reinforcement learning takes over in closed loop, learning from the consequences of its own actions in simulation. 👉 Closing the loop: A release gate blocks a candidate on a statistically significant regression, a broken hard threshold, or a failing sentinel scenario check. Once the model is deployed, its failures on the road get curated into the next round’s evaluation and hard-case datasets. Want to go deeper? Check out Shiyan’s full breakdown here: https://lnkd.in/gxq4VTMP And if you’re headed to LanceDB's #Reverie2026 on Nov. 5th, he’ll be presenting on the main stage at 11:30 AM PT.

End-to-end driving models usually learn from fleet-scale human driving logs. Those logs are expensive to collect to capture the safety-critical long tail. Wei Zhan and his team have developed TerraTransfer++ to remove human demonstrations from every stage of training. This means: • Drive: a policy learns through PPO self-play in our vectorized simulator, TerraZero. • Plan: its rollouts are distilled into a flow-matching trajectory planner. • See: adapters on a frozen DINOv3 backbone align a camera-only student to the teacher, using only paired (image, scene-state) frames. The result: the best of any published method on NAVSIM and HUGSim, the most widely used E2E driving benchmarks Paper: arxiv.org/abs/2610.30319 Project: https://lnkd.in/diWpiq83

Data Engine jumpstarts autonomous capabilities by collecting training data via passive navigation capabilities installed on U.S. Navy ships. The US Navy’s challenge in fielding autonomous and AI-enabled systems is generating high-quality, operationally-representative data to train, test, and validate them. Data Engine is the infrastructure layer that helps collect operational data, curates it, and trains perception models. It then evaluates and deploys those models back to vessels over the air. The result is what one ship learns now reaches the entire fleet – a threat encountered by a single vessel is turned into an improved model and deployed over the air, so every ship can detect it. Future conflicts will be won by nations that can field autonomous capabilities faster and more effectively than their adversaries. That requires infrastructure that can continuously collect data from operational environments, train models on that data, and deploy improved capabilities back to the field continuously. We are building toward a future where each military service operates its own Data Engine, tailored to its domain, and feeding into a common infrastructure. https://lnkd.in/g5F5Cw7F

What does it mean to design a car when software shapes the experience? Our Head of Design and Marketing, Karl Heiselman, toured Stellantis’ campus in Auburn Hills with Chief Design Officer, Ralph Gilles. They talked about how Applied Intuition and Stellantis are working together on STLA SmartCockpit, Stellantis’ in-vehicle infotainment platform, powered by Applied Intuition's Cabin Intelligence, and why design is key to delivering a truly remarkable experience for drivers. As Ralph puts it, the job of designers is “to not forget the human involved.” https://lnkd.in/g-kajtsZ

Last week, Applied Intuition hosted more than 20 CTOs and technology leaders from many of the world’s largest agricultural companies. Agriculture is one of the most important and most underserved frontiers for physical AI. We shared how we’re applying vehicle intelligence proven in other industries to the field, and what stood out most was how strongly our platform, Dana, resonated with the teams who shape how the world grows food. Thank you to the SVG Ventures | THRIVE team for making this possible, and to every leader who joined us. We’re just getting started. If you’re looking to leverage AI in agriculture, we’d love to connect. #Agriculture #PhysicalAI #AutonomousSystems

Learning dexterous manipulation skills via RL poses a challenging exploration problem. Our Chief Scientist, Wei Zhan, and his team resolve it with X-Reset: sampling resets from retargeted human hand-object states. One policy. Diverse objects, from small markers to heavy power drills. Zero real world robot data.