Pinto: academics despair as Astra zero-shots robotics
Lerrel Pinto says Astra, Fable and Muse zero-shotting robotics benchmarks has academics in deep despair, Ken Goldberg calls the shift thrilling, and Chen Tessler says he has no idea how to reproduce the Astra posts.
Read the whole week in one goSkild says $100M ARR, Agility books $1.8M
The big one
Lerrel Pinto says academics despair as Astra, Fable, Muse zero-shot robotics benchmarks
Lerrel Pinto says academics have spent the past week in deep despair because Astra, Fable, and Muse are zero-shotting (solving with no task-specific training) benchmarks in robotics and world models, and calls it what a step jump in progress looks like. Ken Goldberg answered that it is also thrilling to watch a paradigm shift in robotics, and Chris Paxton said the soul searching is happening outside academia too. Paxton also says he keeps finding problems Astra cannot solve, and Chen Tessler compared the Astra posts to pizzas that look amazing but that he has no idea how to reproduce.
Uneasy pill to swallow, but this is what step jumps in progress looks like.
But also thrilling to witness a paradigm shift in robotics.
They look amazing, but I have no idea how to reproduce them.
Out in the world
UBTECH commissions 10,000-unit Walker S factory; Paxton says robots only do kitting
@humanoidsdaily reports UBTECH has commissioned a 10,000-unit-scale Super Smart Factory in Liuzhou, Guangxi, engineered for a 10-minute takt time (the interval between finished units coming off the line). The 14,000 square meter facility uses a flexible, non-conveyor assembly line where mobile manipulators and AGVs (automated guided vehicles, wheeled carts that drive themselves) handle kitting and logistics to build the Walker S series. Chris Paxton's take: the robots are kitting, doing easy stuff, and bringing parts to human workers.
the reason you know we aren't heading for hard takeoff yet, is that the robots are kitting and doing easy stuff and bringing things to human workers
Teleop and other arguments
Humanoid Scott: Musk's Optimus V3 'frame by frame' line is four months old
@polymarketmoney posted that Elon Musk revealed Tesla is holding back Optimus V3 because competitors would analyze it frame by frame and copy it before production. Humanoid Scott says Musk said this on April 22, 2026, during Tesla's Q1 2026 earnings call, and that it is not news to anyone following the space. He also agreed with Rob Forth that Musk's usual line, that prototypes are the easy part, no longer holds for humanoids, echoing his own earlier argument that humanoid prototypes are hard and that V3 is not being shown because the bar for demos is beyond its capabilities, not because of piracy.
This is not news to anyone following the space
In the past Musk's narrative has always been that prototypes were the easy part. With humanoid robots even that's no longer true.
It’s not a scaling problem, is a prototype problem
Also on the timeline
Simon Kalouche posts a Unitree G1 humanoid teardown
Simon Kalouche posted a teardown of the Unitree G1 humanoid, with a full-length video linked. Worth a look if you want to see what is inside.
Paxton calls zero-shot sim-to-real for Asimov an important achievement
Chris Paxton called zero-shot sim-to-real (a policy trained in simulation running on the real robot with no extra tuning) for Asimov an important achievement. His point: sim-to-real only looked easy on the Unitree G1 because the whole international community was working on that one robot.
Kenneth Cassel: 10x more people building robot dev tools than robots
Kenneth Cassel says every robotics founder he talks to is flooded with outbound from dozens of dev tools, data collection, and infrastructure startups, and estimates at least 10x more people are building software tools for robots than building robotics companies. Simon Kalouche replied that he blocked every company trying to sell him garbage data.
Paxton's blog: self-driving trucks are still chugging along
Chris Paxton wrote a blog post arguing self-driving trucks are still making progress, whatever the vibe on X. In another post he said human workers do a lot more tasks than advertised, even in highly bounded roles, and that this is common across industries.