🚢🚢🚢🚢🚢🚢🚢 Big 🚢 incoming from @neuralmotion. Most robot foundation model labs are sitting on months of teleoperation data that will never train another policy. Wrong hardware revision. Dead on arrival. @aurorafeng_01 went and asked the labs what they actually do about it. The answer, more often than not: nothing. Collect again. She started @neuralmotion to kill that answer. I have had the seat across the table from her at @saturdayrobotic while this thesis took shape, and I think the field currently has cross-embodiment pointed in the wrong direction. Five things worth stealing from her framing 👇 1️⃣ Cross-embodiment is not a problem between companies. It is a problem inside one. Everyone models it as a coordination failure — my robot, your robot, whose schema. Aurora relocates it to two revisions of the same arm, on a roadmap owned by one team, in one building. Same problem, vastly larger surface, and you already have it. 2️⃣ The ceiling is set at collection time. Collect 200 tasks per arm, the policy learns 200. Ask for the 201st, nothing to reach for. No compute and no architecture moves that number — it was fixed the day someone decided what to record. The most consequential decision in the pipeline happens before a GPU turns on. 3️⃣ Co-training helped. It did not lift the ceiling. Open X-Embodiment: 22 robots, ~160,000 tasks, real positive transfer. Her objection is sharper than "not enough" — nobody knows why it works when it works. The field spent its attention on the model side. The data recipe, upstream of pre-training, is close to unexamined. 4️⃣ Retargeting solves the wrong space. Kinematics maps joints across arms. That is configuration space. The gaps that break transfer live in observation space and action space, and joint mapping closes neither. Most people nod at this without noticing it invalidates their current stack. 5️⃣ Human video is not a shortcut until you have a robot reference. Hands cover every task on earth, none of it grounded in what a robot doing that task should look like. Everyone scraping more egocentric video has the order backwards. Learn task→embodiment first, and human video stops being a modality and becomes one more body. Aurora and I are from the same hometown — beautiful @WuxiCity, and we reunited on the other side of the world. I initiated @saturdayrobotic; Aurora has been my co-host since Day 1. We ran the first session together, and grew it to ~3,000 researchers in a few months. I’ve driven the weekly programming; Aurora help built the community and the founding team behind @neuralmotion. Proud to see what started in our @saturdayrobotic room helps grow a company day by day. 🚢 @saturdayrobotic warmly congratulates @neuralmotion 🚢 Big ship coming. 🚢🚢🚢🚢🚢🚢🚢 #Robotics #EmbodiedAI #PhysicalAI #RobotLearning #CrossEmbodiment #BayAreaRobotics