Outsiders test GPT-6 Astra: typing yes, block stack no
Researchers are running their own tasks on GPT-6 Astra and finding ones it fails, ETH Zurich's humanoid swings across monkey bars on raw lidar, and Skild AI clarifies its $100M ARR rests on multi-year contracts.
The big stuff
Outsiders test GPT-6 Astra: it learns to type, fails a three-block stack
Andrew Lyubovsky says he still could not get GPT-6 Astra to stack three blocks with a ducky on top, and put up $100 for the first person who does, even after finding that pointing the model at the robot's unused wrist camera improves performance. Kaifeng Zhang asked Astra to express itself on a keyboard and says it took 40 minutes, several accidental long presses and a lot of backspace before it learned to type, in a video shown at 20x speed. On the plus side, Axel showed Astra doing in-context learning (picking up a task from examples, with no retraining) on mobile manipulation in a different environment, camera angle and layout with no text prompt, choosing between end-effector and joint space on its own. Jiafei Duan borrowed @chooi_jeq's code and ran an open-loop rollout (feeding the model a known trajectory and watching what it predicts) on a MolmoAct2 sim trajectory to test whether Astra was trained on robot data, which he says is the only way to trace a closed-source model's training data.
We still couldn’t get it to stack 3 blocks with a ducky on top. $100 to the first person who does it.
40 minutes, several accidental long presses, and a lot of backspace later, it learned how to type.
everyone still astra-posting and yet i seem to keep finding problems it can't solve. curious
Demos and launches
ETH Zurich humanoid swings across monkey bars reading raw lidar, no map
A humanoid from the Legged Robotics Lab at ETH Zurich jumps up to a set of monkey bars, swings across and drops to a landing, per Lukas Ziegler's writeup of the paper. Most legged robots build a terrain map first, and Ziegler points out a heightfield throws away exactly the thin overhanging geometry a bar is made of, so this policy reads the raw scan from a head-mounted solid-state lidar with no reconstruction step. An attention-based encoder picks out which sparse returns matter and a GRU (a small recurrent memory) keeps track of the bar once it slides out of the sensor's view. Training used separate privileged teacher policies for jumping up, brachiating (swinging hand over hand) and jumping down, then distilled them into one student with DAgger (the teacher corrects the student's mistakes), a critic warm-up and PPO (a reinforcement learning algorithm).
Most legged robots build a terrain map first, and a heightfield is the wrong representation for a horizontal bar with air above and below it. Thin overhanging geometry is exactly what a map throws away.
The detail I enjoyed most sits in the sim-to-real section. Alongside LiDAR noise, they modeled battery voltage sag and actuator thermal limits.
Deployments and money
Skild AI's $100M ARR is based on multi-year contracts, Pathak clarifies
Skild AI says it hit 100M ARR within 10 months of starting deployments, across factory lines, construction sites, kitchens and data centers. The new detail: Humanoid Scott says Deepak Pathak, on a visit to Skild HQ, personally verified the current ARR numbers and added a clarification that the figure is annual recurring revenue based on multi-year contracts. Chris Paxton called it good details, and Mattie Fairchild used the moment to pitch a six-week bootcamp at Mission Robotics for 10 funded teams covering deployment, operations, financing and insurance.
@Rewkang and myself visited @SkildAI HQ this evening and @deepakpathak personally verified the current ARR numbers with us and added this clarification to his earlier article
Might fuck around and make a six-week deployment accelerator that brings 10 funded teams to Mission Robotics for a pressure cooker bootcamp covering deployment, operations, financing, insurance, and emerging RaaS models
Quick hits
Chris Paxton reposts his guide to faking a robotics paper before ICRA deadline
With the ICRA deadline a few days away, Chris Paxton reposted his tongue-in-cheek guide to faking a robotics paper: keep your model secret so nobody can run comparisons, control the demo scene, never show failures, pick weak baselines and cherry-pick benchmarks. The one thing a faker must never do, he says, is what Physical Intelligence and NVIDIA do: open-source the model so anyone can benchmark it.
Teradyne sues cobot maker JAKA, its second European case against a Chinese firm
Teradyne Robotics has filed a patent case against JAKA's German arm in Copenhagen, per a report Bill Hughes quoted, its second major European case against a Chinese robotics company this year after suing Elite Robots' German subsidiary over alleged copyright infringement in February 2026. Humanoid Scott marked the occasion with a Beowulf-style poem about cobot kings clashing in parchment and plea.
Léo: count motors, not joints, robot DoF numbers are mostly marketing
Léo argues that the DoF (degrees of freedom, the independent ways a machine can move) number in robot launches does the same job a teraflops figure does, since coupled linkages let one motor bend several joints: Inspire's popular hand moves twelve joints with six actuators, so the honest count is six. Humanoid Scott agreed and says he prefers the term degrees of control since degrees of freedom has lost its meaning.