Astra drives a robot arm from one human video, Xiao says
Wenli Xiao says GPT-6 Astra copied a human demo onto a robot arm on the first pass, Ken Goldberg says the frontier is now harness design, and Markus Wulfmeier's benchmarks say Astra still has ten times the error rate of his team's internal models. Also: Skild AI claims $100M ARR, Figure robots wander HQ, and Maven leaves stealth with $100M.
The big stuff
Wenli Xiao says GPT-6 Astra drove a robot arm from one human video
Wenli Xiao says his team dropped a recording of a human doing a novel task into the Codex app, prompted it to drive a robot arm the same way, and it worked on the first pass. He calls this physical ICL (in-context learning, where the model picks up a task from an example in the prompt with no retraining). Ken Goldberg says agentic robotics is the missing link between model-based and model-free methods and that, to fix latency, it is now generating executable ROS2 code (a common robot software framework). Separately, @huxiao93612565 says Astra wrote an entire real2sim pipeline (turning a real video into a physics simulation), including hand tracking, IK retargeting (mapping the human hand motion onto the robot hand) to a 44-DOF (degrees of freedom, roughly the number of joints) dexterous hand, and grasp refinement, from one video of a human hand with zero code from the user, and is open-sourcing it.
GPT-6 Astra just does physical ICL out of the box.
To address latency AR is now generating executable ROS2 code.
A large gap still remains to our internal models (x10 lower error rates) but I'm excited to see the recent advances.
Deployments and money
Skild AI says it hit $100M revenue run-rate 10 months into deployments
Skild AI says it reached $100M ARR (annual recurring revenue run-rate) within 10 months of starting deployments, and lists factory lines, construction sites, kitchens and data centers, with robots cleaning, welding, building and cooking. Skild builds the AI brain for robots, software that Ziegler says lets them learn new tasks across different bodies and environments, not the hardware. Ziegler quotes Deepak Pathak calling it a new chapter, moving from an era of demos to an era of deployments.
We just hit 100M ARR within 10 months of starting deployments.
This signifies a new chapter for robotics, moving from an era of demos to an era of deployments.
I just think data visualizations can be deceptive.
Discourse and drama
Scoble says Figure robots roam HQ unwatched; Erik Nieves streams double Figure's pick rate
Robert Scoble says Figure's San Jose HQ has robots walking around with no one watching them, and Biti8 says multiple robots move around the campus without someone controlling every step. Adcock's reply was that the Figure campus is year 2050. On the other side, Porter says Nieves, who he describes as having 35 years of industrial robot deployments, two billion picks and 200 plus deployments, answered Figure's 1,200 per hour parcel induction stream (placing packages one at a time onto a sorter) with his own stream at 2,200 to 2,400 on bulk flow with a full SKU mix, framed as what real entitlement looks like (industrial automation talk for the rate a line is expected to sustain). The Deployed episode is out.
At its San Jose headquarters it has a bunch of robots walking around with no nerds watching them.
max AGI is a humanoid robot that can do everything a human can
The gap between a demo and an entitled workload is the whole conversation.
Quick hits
Maven Robotics leaves stealth with $100M and a dual-arm palletizing robot
RoboStrategy led Maven's $100M Series A, with LocalGlobe, Vine Ventures and XTX Markets also in, for wheeled dual-arm robots that Lukas Ziegler says move at 10 mph with a 30 kg payload for mixed palletizing (stacking assorted goods onto pallets), with 250 third-gen robots planned and a fourth-gen platform in design. Ziegler also relays CEO Hamza Derbas saying bipedal humanoids make zero sense in warehouses because they are complex, unreliable and add cost.
Kevin Zakka's mjbatch trains a quadruped on a laptop in one minute
Kevin Zakka released mjbatch, a tool that steps thousands of MuJoCo (a physics simulator) environments in parallel on a CPU from Python. He says a Go1 quadruped learned joystick following in 1,024 parallel environments in one minute on his five-year-old M1 MacBook, and a G1 humanoid backflip was optimized in 40 seconds on the same laptop.
Kat Scott cheers McKinsey report calling humanoids a media distraction
Kat Scott says she saw a McKinsey report on Physical AI that characterized humanoid robots as a media distraction and forecast they would capture only 2% of the potential Physical AI market by 2045, and she agrees. Only her description of the report is in the sources, the same day Maven's CEO said humanoids make zero sense in warehouses.
Jeannette Bohg wants robotics conferences to be all workshops
Agreeing with Michael Black's post on academic publishing, Jeannette Bohg says she has been bored by CoRL's (a robot learning conference) main track for years, already knows 90% of the work, and proposes making conferences all workshops and panels, while noting publications remain the measure of academic success. In a follow-up to his own post, Black said academia lacks the resources even to evaluate top models and suggested companies give compute tokens back to academia as a kind of tax.