A robot hand with 50 DoF turned out easier to teach than a gripper with 1. @JasonMa from @DynaRobotics at @RoboticsCtr last Saturday, ~200 people: after 1M hours of human-video pretraining, the bottle-cap task took 13 min of robot data. Some gripper tasks took hours. Models are ready. They're waiting on the hand. Thanks for @junfanzhu98 and @aurorafeng_01 from @saturdayrobotic for cohosting. – at San Francisco, CA

The robots that pass every eval are still killing their companies. Just slower. Last week's post on RL² got one piece of pushback I can't stop thinking about. It's the real reason robotics fleets die. The pushback: A robot that passes every eval is still a cost center—CapEx, maintenance, operators, downtime. Most fleets don't die in the lab. They die at the pilot stage, when the unit economics never close. That's the part that never makes the pitch deck: even a model that clears every eval still has to survive a business model. And usually, it's the business model that ends the pilot. It also breaks the standard engineering flywheel: → Deploy → Capture failures → Evaluate → Collect targeted data → Redeploy That loop only keeps running if each cycle costs less than what it buys you. Where I've seen it stall: Hardware bought before there's enough throughput to justify owning it. Data collected that doesn't target your actual edge cases. No real gate to tell the difference between a pilot that's improving and one that's quietly bleeding money. None of that is a modeling problem. It's a sequencing and cost-allocation problem—and it's why teams with genuinely good models still stall at pilot. That's the real reason we built one stack: hardware you don't have to own, data aimed at your specific failure modes, and an eval gate that tells you when you're actually ready to scale. The goal isn't just to close the sim-to-real gap. It's to make the loop cheap enough to survive deployment. The bar isn't "did the sim-to-real gap close." The bar is "did it close before the pilot ran out of runway." If your pilot works but doesn't pencil, that's the conversation worth having at our event: Robotics Center RL² Takes the Field Event this Sunday, July 19th.

Everyone in the Bay is building a robot company. Almost nobody is talking about the real bottleneck. After 2 years running the Robotics Center of Silicon Valley, seeing hundreds of frontier labs and stealth startups, one pattern is undeniable: The blocker isn't model capability. It’s the sim-to-real loop. 1. Simulation ≠ Hardware Reality A VLA trained in Isaac Sim or MuJoCo hits a wall when it encounters real-world actuator limits: thermal throttling, joint backlash, and sensor drift. Even two units of the exact same robot model have completely different dynamics. 2. Teleop data is inherently off-policy Human demonstrations bootstrap your model, but your model isn't human. That distribution gap doesn't disappear with more of the same data. It only closes through real-world rollouts, failure capture, and targeted data collection. 3. "Working" is undefined until deployment Teams optimize for lab success rates, then fail on real-world edge cases: a wrinkled shirt, 4 PM glare, or a pallet 3cm off spec. Without a real-world evaluation harness, you are just guessing. 4. The solution? Treat RL² as your core loop Internally, we call this Reinforcement Learning in Real Life. Deploy ➔ capture failures ➔ evaluate against real criteria ➔ collect targeted data ➔ redeploy. Continuous learning on real hardware. 5. The Takeaway The winners in Physical AI won't just have the biggest models. They'll be the teams that can close the real-world learning loop the fastest. Closing the sim-to-real gap is a 2-year infrastructure problem. That’s why we built our stack at the Robotics Center. 6. Come hang out If you're hitting these exact walls, let’s compare notes. We host robotics builders at 90 Welsh St (SF) every Friday. Drop by.

We just unboxed the WUJI Glove — one of the first in the world to get it. 60g tactile data glove with 800Hz IMU, 526-point pressure sensing, and under 2mm motion accuracy. Built for robot teleoperation and imitation learning data collection. Get it today: [email protected] #Robotics #PhysicalAI #EmbodiedAI

World's first WUJI Hand unboxing — and we brought in a real user to tell us exactly what they think. 20 active DOFs. Get yours → [email protected] #WujiHand #Robotics #EmbodiedAI #Unboxing