Machenta's two-minute factory, and the Astra fight
Edward Johns says Machenta taught a three-robot packaging line in two minutes with a model trained entirely in simulation. Meanwhile the field spent the week arguing about whether OpenAI's Astra is robotics' GPT-3 moment when almost nobody owns a robot to test it on.
Shown this week
Machenta's one-demo learning held up across a week of clips
Machenta's claim survived a week of follow-up videos: one demonstration, no further training, and the same model driving three different bimanual robots on a packaging line. Edward Johns says the line took two minutes to teach, the model was trained 100% in simulation, and it had never seen those objects. He also posted a steerability test, teaching three different interactions with the same objects just by changing the demonstration.
Normally, this would take weeks of system integration, but it took us two minutes to teach.
Two new papers say robot manipulation needs force, not just motion
Contact and force became a manipulation theme this week across two separate releases. Cunxi Dai's CARD tracks real-world contact and forces alongside kinematics, running at 10 Hz with no object models and no simulator. Haonan Chen's VisualForce reads gripper deformation to handle tasks like berries and plug insertion with no force or tactile sensors at deployment.
Can you believe these are open-loop rollouts?
Out in the world
Tutor raises $75M and targets 1,000 robots; Figure ships F.03 off the line
Tutor raised a $75M Series B led by a16z and says it is on track for more than 1,000 robots under contract by the end of 2027. Josh Gruenstein says Tutor is in almost 4x as many sites as a year ago, and nearly a third of this year's deployments went to existing customers coming back.
we’re on track to have more than 1,000 robots under contract by the end of 2027
Teleop and other arguments
The field split over whether Astra is robotics' GPT-3 moment
Freda Duan's thread arguing Astra is robotics' GPT-3 moment turned into the week's main argument. Brad Porter says Astra generalizes better than anything he has seen publicly but is not outperforming his own lab's general model, and that distilling it down to production performance may not be easy; he still expects general models industry-wide in two to three quarters.
Not everyone has a robot to test Astra.
Also on the timeline
Meta reportedly built a humanoid called Max with Fourier
Meta reportedly built a prototype humanoid named Max with Chinese startup Fourier Robotics in about three months, weighing roughly 48 kg.
Foxconn is not building its own robot brain
Foxconn's Leo Guo told Lukas Ziegler the company runs NVIDIA's GR00T 2.0 rather than training its own foundation model. Its edge is post-training on real production line data.
Frequently asked questions
What is Machenta and why does it matter?
Machenta is a robot learning system that can control different bimanual robots with just one demonstration and no further training, using a model trained entirely in simulation. The key claim is that it works on real robots without any real-world training data.
Why are robots adding force sensing to manipulation tasks?
Two new research papers show that tracking contact and forces alongside motion helps robots handle delicate tasks like picking berries or inserting plugs. Some methods can read force information from gripper deformation without needing additional sensors at deployment.
Is Astra a major breakthrough for robot AI?
The robotics field disagrees. Some researchers say Astra is a major milestone similar to GPT-3 in AI, while others argue it does not outperform their own lab models. Nobody outside the labs has tested Astra on real robot hardware, so both sides are debating results that are not publicly verified.
How many robots does Tutor expect to deploy?
Tutor says it is on track for more than 1,000 robots under contract by the end of 2027. These robots are controlled by human tutors in the loop rather than operating fully autonomously.
Is Foxconn building its own robot AI model?
No. Foxconn runs NVIDIA's GR00T 2.0 foundation model rather than training its own. The company's competitive advantage comes from adapting the model using real production line data.