everyone assumes robotics is bottlenecked by data. collect enough trajectories and eventually robots just get smart, right? @axisrobotics is quietly proving that's not the full picture. here's the part that got me: a small, carefully picked dataset actually beat using the entire data pool. their Sim2Real numbers moved from around 22% to 52% in the process. so more data wasn't the unlock. better data was. then there's the experts angle, and this is the part I can't stop thinking about. instead of training one giant model to do everything, they train narrow "Experts" on single skills. one result showed near 100% task success from an expert that cost somewhere between $5 and $10 in compute. that's a small number for that kind of result. this isn't just a data company anymore. it's starting to look like a skill factory, small pieces trained well, then combined for bigger tasks later. genuinely curious where this goes once they start combining multiple experts into one robot. still in the network for this one :) join Axis Robotics: s.kaito.ai/E6hkB0B

most creators are gonna miss this window. @axisrobotics just opened epoch 2 of their creator program. sep 19 to oct 18, thats it. ive been creating around their robotics and physical ai work and the momentum is real right now. still time to jump in and earn a reward. who else is in this round?

you know a project is good when the art alone makes people stop that's exactly what happened with @fomoater for me and my community Overthinkers has been coming up a lot lately, people genuinely curious about it @breyonchain would love to team up if you're open, happy to bring some solid people into the community 🚀

9 months. that's how long it took @axisrobotics to build a Physical AI data engine that makes 100,000+ data trajectories every single day. now the $AXIS community sale is oversubscribed, with $1M+ committed already. but it's still open until Sep 28, 1:00 PM SGT. (11 AM BD time) you get a pro-rata share. whatever isn't accepted comes back to you in USDC. are you in or just watching? 👇

all my friends already got their eligibility on @zaddrnet and i'm the only one still stuck on pending i've been refreshing all day with a little hope each time. and every time, nothing. @zaddrnet please notice me this time 😭

The Axis badge claim just went live and barely anyone noticed. Go to Axis Hub Open your profile Follow their account @axisrobotics Connect your X Badge claims itself automatically No extra steps, no forms. Theres also a Kaito Katalyst reward pool live right now. Quarter percent of total $AXIS supply going to creators. • Top 500 yappers split it, early spots getting double the smaller ones. • Fully unlocked at TGE. • Epoch runs August 19 to September 18. Are you claiming yours today or waiting it out?

Your robot doesnt need more data. It needs better questions. Been digging into @axisrobotics work lately and this part actually stuck with me. Most teams just keep piling on more training examples and hoping the model magically gets better. Axis is doing something different. They let the model basically point at itself and go, here, this is where I keep messing up. So instead of more data everywhere, you add data exactly where the failures are happening. Train it, catch the gap, fix that one spot, run it again. And its not just one type of data either. Simulation, human demos, real robot runs, mixed together depending on what the problem actually calls for. The Booster project is a solid example, simulated environments paired with real demonstrations. They're also linked up with teams like Feagine and Manycore on the physical AI side. Honestly the dataset size stopped being the interesting part to me a while ago. What gets me is a model that can basically tell you what it still needs to learn before you even ask. Would you trust a model more if it could flag its own blind spots, or does that make you more skeptical? Join Now s.kaito.ai/E6hkB0B

Spent an hour comparing @axisrobotics's numbers to what's actually verifiable on Base. The gap between "crypto project with a robotics story" and "robotics project that happens to use crypto" is bigger than I expected here. 5M+ trajectories is on-chain, not a landing page claim. 200K contributors is real too. But here's what most threads skip that's contributors, not buyers. Two different funnels. Only one of them decides whether $AXIS actually has demand behind it in three months. Sale terms, for anyone tracking: $0.10/token $100M FDV $1M raise $100–$100K per verified entity (KYC via Sonar, not just a wallet connect) 10% at TGE → 6mo cliff → linear over next 6mo Window: Sept 21–28 Settlement: ~Sept 30 Worth noting, it's Reg S. US persons and a list of other jurisdictions are excluded outright. That cuts the buyer pool more than people assume when they're doing "$1M raise, uncapped upside" math in their head. $100M FDV on a $12M seed with real Coinbase-adjacent traction? Not the part I'd push back on. The part I'd actually dig into before committing capital does 90% of supply unlocking linearly post-cliff get absorbed by real usage, or just hit the market at month 7 regardless of demand. Not a call either way. @axisrobotics is building something with actual on-chain receipts behind it just don't think the tokenomics are as simple as the sale page makes them look.

Robotics doesn't have a model problem. It has a data problem. @axisrobotics is quietly fixing that. No hardware. Just a browser. You control a simulated robot arm and every move becomes real training data for tomorrow's Physical AI. The numbers right now say it all 👇 🔹 5,210,492 total trajectories 🔹 205,102 contributors 🔹 4,796,727 on-chain transactions 🔹 709 new tasks in the last 7 days alone And it's not just scale, it's quality. Their Sim Dataset V1 pushed π0.5's success rate on LIBERO-Plus from 83.9% → 88.8%, beating an equal amount of RoboCasa365 data by 31 points. If robotics gets its "GPT moment," it won't come from a bigger model first. It'll come from someone solving the data supply chain. Axis is building that supply chain, one trajectory at a time. 🔵 Are you contributing yet? #PhysicalAI #Robotics #Base

I've been staring at this @axisrobotics thing for a bit and its stuck in my head for a weird reason. Everyone talks about robots learning to grab or fold or sort stuff. Nobody talks about who decides what they even practice on first. Sounds small until you sit with it. Every task a robot learns from starts as someone deciding an object, a room, an instruction, and honestly like five different ways of phrasing that instruction. Now do that a few thousand times by hand. I tried picturing someone doing that manually and just gave up halfway through the thought. Axis built something called AxisTaskGen for this. You type a plain sentence, "put the ketchup into the basket" is literally their own example, hand it whatever objects you have, and it builds out a full task layout ready for simulation on its own. A person still checks the layout before it gets used though, which I wasnt expecting honestly. Makes sense though, an AI generated setup can look fine and still be wrong in some way you'd only catch by eye. Not gonna lie, I care less about the robot arm now and more about whoever or whatever decided what it got to practice on in the first place. Is task data the actual bottleneck in robotics or is that just me @KaitoAI

everyone is chasing airdrops but sleeping on the easy ones. @axisrobotics just opened a @BinanceWallet campaign and its live right now. → connect your wallet to Axis Hub → find the campaign, → complete the available tasks → sign when it asks. thats it. no grinding, no gas wars, just showing up. are you actually doing these or scrolling past every time?