CBoston Dynamics has Atlas doing handstands, and Figure says its new robot brain finished half its tasks in thirty homes it had never seen.
MHalf its tasks? That number is doing a lot of work.
CFifty-six percent of full tasks. No partial credit, so almost doesn't count.
MWhich sounds impressive until you ask what the other forty-four percent was.
CNine percent without the new human-video pretraining. The jump is real.
MNine percent is a robot that folds one towel and knocks over a lamp.
CElinaNotes on X posted, "The humanoid robotics race isn't about demos anymore." She says it's about the brain, manufacturing scale, and reliability.
MThat's three different bets, and only one of them is proven.
CFigure is betting on the brain, and Figure 03 is the first one built for a factory line, not a demo stage.
MBuilt, maybe. But Finance Spot's post on X has Tesla, Figure, and 1X all ramping production. Which part is confirmed?
CFigure's BotQ line is up and running, with big production targets ahead.
MTwelve thousand a year is a real number. A hundred thousand is a slide deck.
CThey vertically integrated actuators, batteries, sensors. They're not just assembling parts.
MBut that doesn't answer whether anyone needs twelve thousand humanoids walking around a factory.
CIt answers whether they can make them. That's the whole bottleneck!
MNo, the bottleneck is the brain. A handstand is a party trick. Try stocking shelves in an unfamiliar warehouse.
CThat's exactly what Helix is for. It uses video of humans doing tasks, then the robot mimics them in new homes.
MAnd the fifty-six percent was in thirty unseen homes. Thirty is a tiny sample, and the tasks were tidying, folding towels, making beds.
CWhich are exactly the kind of unstructured tasks robots have failed at for a decade.
MBut Faris Capital says the strict rule means no partial credit. So a robot that almost folds a towel scores zero.
CGood. That's the only number you can trust. Soft metrics are how we got robot demos that only work in one room.
MI'll give you that. But a strict fifty-six percent also means it's failing nearly half the time in a real home.
CIt's a progression. Not a finish line.
MI want to see the robot that can find the missing sock. Then I'll be impressed.
CWe can argue about socks after Figure ships a robot that doesn't fall over on carpet.
MBoston Dynamics Atlas did a handstand. That's nice. But it was reinforcement learning in a lab, with a safety tether presumably out of frame.
CActually, the video from AIOSTIK on X shows Atlas walking, running, crawling, getting up from falls. No hand rails.
MAnd how many times did it fall before the camera rolled? Reinforcement learning means thousands of attempts.
CThat's not a flaw. That's how learning works. The robot earned that handstand.
MYou're romanticizing a machine doing calisthenics.
CI'm not. I'm saying the physical part is solved enough to start scaling.
MSolved? We still don't have a humanoid that can reliably open a door it hasn't seen.
CBut Figure reported the latest Helix model with human-video pretraining finished fifty-six percent of full tasks in thirty unseen homes. That's a door, a towel, a bed.
MIn homes, not factories. Last I checked, factories are full of conveyors, forklifts, and workers who move fast.
CRight. So why not start in places where labor is short and turnover is high?
MBecause a robot that drops a box once an hour can shut down an entire line. A human doesn't.
CBut a human gets injured doing the same lift twenty times an hour. The robot doesn't.
MThat's the pitch. But no factory pays a robot to be a safety poster. It has to hit cycle times.
CFigure claims twelve thousand a year. If even half of those go into factories, you'll see the data.
MI'll wait for third-party data, not Figure's own blog.
CThen you'll wait until the robots are already installed. That's not skepticism, that's missing the boat.
MNo, that's reading the fine print. One robotics tracker says these percentages across Figure, NVIDIA, Physical Intelligence, DeepMind aren't comparable. Different tests, different hardware.
CSo compare on the same task, then. The fifty-six percent was full task completion, no partial credit. That's already stricter than most demos.
MStrict but still cherry-picked. Thirty homes, selected? They didn't say randomized.
CIt's a research paper, not a clinical trial.
MExactly my point. It's not a production-grade result.
CI'll take a research result over a concept video. Finance Spot's post mentions Tesla Optimus Gen 3 production ramp. That part I don't trust.
MAgreed, and that's the problem. The hype blends confirmed Figure numbers with unverified Tesla and 1X ramps.
CSo let's separate: Figure is real, Atlas is real. The rest is aspirational.
MAtlas is real and can do a handstand. My Roomba can't do a handstand, but it cleans my floor every day.
CYour Roomba has one job. These robots are being built for dozens.
MAnd they're doing each of them inconsistently. That's not a workforce, that's a circus act.
CIt's an early industry. You're judging a five-year-old by how well she performs surgery.
MNo, I'm judging a two-billion-dollar industry by the claims it makes.
CThen hold it to the claims that are confirmed. Figure 03 exists. BotQ is building. Atlas moves without a script.
MI'll hold it to the spreadsheet. Show me quarterly production and customer retention, then talk about scale.
CThe spreadsheet is the slowest thing to update. You want the number before the company has it.
MNo, I want the number before it's in a post that says "continues to accelerate."
CFine. We'll revisit when a factory manager posts a video of eight Figure robots on a line at once.
MI'd settle for one robot folding a fitted sheet. Actually, nobody can fold a fitted sheet.
CThat's the real breakthrough. If Figure can fold a fitted sheet, I'll buy one.