Zero-shot dexterity, and 300+ Cybercabs in Austin
A robotics startup says it went from zero to $100M in annual recurring revenue in six months, and nobody outside has checked a single number. Meanwhile two labs taught robots to thread nuts and catch boxes with no real-robot training data at all.
The big one
MicroAGI claims $100M ARR six months after launching Shift
Bercan Kilic says the company went from zero to $100M in annual recurring revenue in the six months since launching Shift. He also says it hosted Europe's first Physical AI summit. Animesh Garg amplified the number and said the mission is a million robots in the real world.
0 to $100M ARR in 6 months is an absurd benchmark.
Shown this week
Two labs did contact-rich manipulation with zero real-robot data
Two separate teams showed precise robot skills learned without a single real-world demonstration. Octi Zhang's Success-Guided Sampling meshes gears, threads nuts and inserts pegs zero-shot from RGB cameras, with no demos and no per-task reward tuning, and he says every video runs at 1x. Workhorse taught a Unitree G1 to topple and climb a suitcase and catch a thrown box from human demos, no teleop (no human driving it).
Everyone says dexterous manipulation needs real-world data. These policies were trained with none.
Goose tidies a house for a toddler using off-the-shelf open models
Humphrey Hu showed a robot called Goose tidying a home after a user asked it to make things safe for a visiting one-year-old. He says it runs Qwen3-8B for the slow reasoning on an RTX 5090 and NVIDIA's GR00T n1.7 for the fast control onboard a Jetson AGX Thor, and that this used to take a team of researchers and a billion dollars.
my timeline for home robots has definitely moved up
Out in the world
Tesla says Austin now has 300+ Cybercabs, 8x in 30 days
Tesla's Robotaxi account says more than 300 Cybercabs are now serving Austin, an eight-fold increase in the 30 days since launch. Ashok Elluswamy says the cabs are everywhere in Austin and Dallas is next. In the same week Elon Musk said a digital Optimus can play about halfway through the Diablo campaign just by looking at the screen like a human.
Teleop and other arguments
Adcock says Figure lifts well past 50 lbs, Paxton backs off his chart
Brett Adcock says Figure can move items much heavier than 50 lbs even with full loco-manipulation, contradicting the chart Chris Paxton used to argue it cannot. Paxton had said neither Figure nor Optimus is strong enough for the roughly 50 lb payload most human work needs, which is the standard OSHA limit for manual material handling.
We can do much heavier than 50lbs even with full locomanip
Also on the timeline
QF3 trains humanoids from scratch with one off-policy update
Chung Min Kim released QF3, which he says trains humanoids from scratch, fine-tunes VLAs and even fine-tunes image models with the same update rule.
Ultra raises $62M to fix robot deployment, not demos
Ultra announced $62M in funding, including a $50M Series A led by Framework, arguing deployment is the real bottleneck. Ryan Julian called it the end of the demo era.
Frequently asked questions
What is a VLA and why does it matter for robots?
A VLA is a vision language model, which processes visual information and language to help robots understand and perform tasks. The story mentions debates about whether VLAs are even necessary for basic robot tasks like picking and placing objects.
How did robots learn manipulation skills without real-world practice?
Two labs used a technique called resampling practice around states with moderate success rates, which works without real-world demonstrations or human teleoperations (direct human control). One team's method, called Success-Guided Sampling, could mesh gears and thread nuts from just camera images with no human examples needed.
What is the difference between robot demos and real deployment?
Demos are controlled tests showing what a robot can do in specific scenarios, while deployment means the robot actually working in real-world conditions solving real problems. One company raised funding arguing that deployment, not demos, is the real bottleneck holding back robot progress.
How much weight can humanoid robots currently lift?
There is disagreement about robot payload capacity. One estimate suggested robots cannot lift the roughly 50 pound standard that human workers typically handle, but a robot company founder disputed this claim, though neither side released video evidence of their claims.