GPT-6 Astra 46%, MolmoAct2 12%: Paxton blames the baseline
Jay Chooi's thread puts GPT-6 Astra nearly four to one ahead of MolmoAct2 on two-armed tasks, and Chris Paxton says that says more about weak, contact-light robotics baselines than about Astra.
The big stuff
GPT-6 Astra beats MolmoAct2 46% to 12% on two-armed tasks, thread says
Jay Chooi (@chooi_jeq) posted a thread saying GPT-6 Astra scored 46% against 12% for MolmoAct2, which the thread calls a state-of-the-art VLA (a model that turns camera images and text into robot actions), across 200 trials on five bimanual (two-armed) tasks, 3.9x higher. Chris Paxton called that concerning for anyone building robot foundation models rather than deployments, then said the gap is mostly because most robotics baselines are contact light, and many pick and place tasks work fine with SAM (Segment Anything, a vision model) plus a very light language model. Separately, Yanjie Ze posted that GPT6 Astra solved a Rubik's Cube with robot hands.
GPT-6 Astra scores 46% vs 12% for MolmoAct2, a state-of-the-art robotics VLA, across 200 trials on five bimanual tasks. That's 3.9x higher.
If you are a robot foundation model guy and not a deployment guy this must be really concerning
most of the robotics baseline are not so good -- theyre very contact light. A lot of pick and place tasks you can do pretty well with segment anything (SAM) and a very light language model.
Demos and launches
@enactic_ai's OpenArm open-sources a humanoid: CAD, firmware, simulation, and teleop
Lukas Ziegler profiled OpenArm, an open-source humanoid from @enactic_ai in Tokyo that ships with full CAD, control code, firmware, and simulation tools. The arms are compliant and backdrivable (you can push them by hand and they give), and teleoperation (a human controlling it remotely) comes with force feedback and real-time gravity compensation. It works with MuJoCo and Isaac Sim so policies can be tested in simulation before touching hardware, and it comes as a kit or prebuilt.
The arms are designed to be compliant and backdrivable. Teleoperation is supported, with force feedback and real-time gravity compensation so operators can guide the arm naturally.
Research
Jiafei Duan's LIT trains robots to act first, see second, and they generalize better
Jiafei Duan introduced Latent Interface Training, or LIT, a two-stage recipe for VLAs (models that turn camera images and text into robot actions) and what the paper calls WAMs. Stage one trains the action expert with no images at all, conditioned on language, robot state, and the end-effector pose each action chunk should reach. Stage two adds vision only through a latent interface that is trained to reconstruct that same pose, so the model cannot lean on backgrounds, lighting, or camera angles that happen to correlate with actions. On LIBERO-Plus, LIT lifts π0.5 from 68.97% to 79.67%, MolmoAct2 from 63.62% to 71.92%, FAST-WAM from 51.44% to 60.63%, and ImageWAM from 83.02% to 86.89%.
Learn to act first, then learn how to use vision.
Our takeaway: how vision reaches the action expert matters for generalization.
Quick hits
Brett Adcock says Figure passed 86,000 weekly active users uploading data
Brett Adcock says Figure crossed 86,000 weekly active users uploading data and calls it the largest and most diverse robotics dataset in the world. The site has a live map of uploads, but the post gives no hours, no task breakdown, and nothing about what the data has trained so far.
John Carmack says Jetson Thor has too much memory for real-time robots
John Carmack argued that Jetson Thor, with 128 GB of memory but only 273 GB/s of bandwidth (how fast the chip can read that memory), is over-provisioned with expensive memory for real-time robotics. His math: a model that runs at tens of frames per second cannot use more than about 10 GB of weights, so most of that memory sits idle. He allowed exceptions, like a wide mixture of experts (a model that splits its weights into many specialists and runs only a few per step) or a large planning model running slowly in small time slices, and said more memory always makes development easier. The real story: this is one engineer's take, not a benchmark, and his point is that a cost-optimized robot brain should get by with much less than the flagship ships with.
Dirac says one customer gets 50% more factory throughput, Kalouche says charge more
Fil Aronshtein says one of Dirac's customers, a power system manufacturer supplying data centers, told them Dirac raises factory throughput by 50%, from $30M a month to $45M a month, a $180M a year revenue lift. Simon Kalouche replied that if it is true Dirac needs to charge more. It is a single customer anecdote relayed by the company, with nothing else in the sources to check it against.
AUAR's MicroFactory builds a home's timber structure in about 12 hours, Ziegler reports
Lukas Ziegler profiled Automated Architecture, or AUAR, whose MicroFactory is a compact, deployable robotic cell that turns timber into a home's structural panels and, per the profile, can fabricate the structure of an entire home in around 12 hours with up to 75% less on-site labour. AUAR says panels come out 10 to 30% cheaper than builders pay today, and the profile reports around 300 homes worked on in the past year with projects in Belgium. AUAR owns the cells, deploys and operates them on site, charges per panel, and targets 1,000 MicroFactories producing 200,000 homes a year by 2030.