We train our own foundation models at @VerneRobotics, and scaling them like this would have been almost inconceivable for us before the @ycombinator × @togethercompute partnership. To truly democratize physical AI, we also have to democratize access to compute. We’ve loved working with @togethercompute and are excited to keep scaling together as we push the frontier of physical AI and build foundation models that deliver real value for our customers.

Robotics is not going to be solved by one type of data. @neil_nie_ sits down to share @VerneRobotics's thinking around robot learning across varied data sources. Different stages of robot learning call for different data modalities. Egocentric human video can provide broad priors. Hand-held data collection devices can efficiently capture dexterous manipulation. Teleoperation remains one of the best ways to collect high-quality robot demonstrations for many tasks. As models improve, other sources of data will become increasingly important as well. The challenge isn’t choosing one approach; it’s building a policy learning pipeline that supports all of them. At @VerneRobotics, we've designed our stack around that philosophy: from data collection, to model training, to deployment. Supporting multiple data modalities gives us the flexibility to use the right tool for each stage of learning, iterate more quickly as requirements evolve, and continuously improve robots after deployment. The winning robotics companies won't be defined by a single data collection method. They'll be defined by how well they integrate many of them into one learning system.

1/ QUESTION: what is a century-old industry that out-precisions human laborers by 5x and has reduced manufacturing injuries by >60%? INDUSTRIAL ROBOTS today, industrial robotics is a tiny market. we explore why this gap exists, and why it is finally closing.

One of the hardest problems in robotics is the deployment gap: the last mile from lab demo to real-world deployment. Over the past two weeks, the Verne Van has been rolling through SoCal customer sites to close that gap. With just a few hours of onsite data, the robot adapted to the customer’s workflow and began autonomously packing boxes of fabric bolts. That speed and generalization is made possible because of how we build: vertically integrated physical AI. We own the full autonomy stack: data collection, model training, and deployment with continual learning so we can iterate quickly with customers and turn real-world edge cases into better robot behavior. This is made possible by the relentless work of the Verne Team. It’s been an exciting 2026 for @VerneRobotics, and we cannot wait to share more of what we have been building. If you run a logistics warehouse and are still bottlenecked by manual pick-and-pack, come talk to us. We would love to show how Verne’s dexterous robot arms and powerful mobile bases can help improve your operations!

Want to have a closer look at @ycombinator Summer Batch 2025? Check my latest article for @Forbes. You won’t be disappointed. forbes.com/sites/dariashunin… @stake_jevens @joefioti @alessiapacca @carlosjc__ @michwirantono @Neha_Suresh_M @pederzh @SepandD @adithreddi @grx_xce @neil_nie_ @ethanmgibbs

It's 2025. If you work in hospitality and think it should be automated — it should. Meet Nemo3 by @VerneRobotics: a dexterous, general-purpose hospitality robot. Laundry. Towel folding. Amenity kits. We do the jobs other robots find too hard. Book a demo → vernerobotics.com/work-with-…

Say goodbye to laundry folding! We’re building that future @VerneRobotics. Interested in our robots? DM me or book a pilot today: vernerobotics.com/work-with-…

🤖@VerneRobotics builds AI software for robot arms to learn new skills in hours (their AI model is SOTA), making it easy for any company to adopt automation. Congrats on the launch, @neil_nie_ and @nolitadrip! ycombinator.com/launches/O3T…