Reward AI's OM-1: no teleop, no robot data
Reward AI says its new OM-1 model learned only from humans wearing a sensor glove, never from teleoperation or robot data, and still runs tabletop arms, industrial arms and humanoids at human speed. On the same day, GPT-6 Astra needed two hours and four human interventions to plug in one ethernet cable.
Read the whole week in one goSkild says $100M ARR, Agility books $1.8M
The big one
Reward AI launches OM-1, a model trained on human hands, not teleop
Reward AI introduced OM-1, which it says zero-shot generalizes (works on a new robot with no extra training) to tabletop arms, industrial arms and humanoids, learned directly from human manipulation data with no teleop and no robot data. The company also posted a test of OM-1 unplugging an ethernet cable, a connector that only releases when the latch is pressed precisely. Yunfan Jiang described the stack: a 7 DoF Omnibody Hand wearable, tactile plus proximity plus in-hand cameras, one policy trained in a single stage, and control trained by reinforcement learning in simulation.
No teleop, no on-robot data, and still human-speed execution across arms and humanoids
I saw the system in action at @RewardAI_ and could hardly believe what it was capable of.
very impressive speed! retargeting may need some work however…
Out in the world
Robocurve raises $10M to independently test robot AI claims
Jay Chooi says Robocurve, an independent Public Benefit Corporation, raised a $10M seed led by Initialized to evaluate frontier AI in the physical world. He says its evaluation harness has been downloaded more than 97,000 times in under three months, with researchers from more than 200 institutions signed up to build benchmarks. The harness is open source and the benchmarks are published so anyone can rerun them.
The more third-party evaluators, the better society can understand how fast robotics AI is progressing.
Teleop and other arguments
GPT-6 Astra takes 2 hours and 4 human nudges to plug in a cable
We have covered Astra zero-shotting robotics benchmarks. What is new is a timing number from a hands-on run: Kaifeng Zhang says Astra plugged in an ethernet cable after two hours and four human prompt interventions, and argues speed is the real problem for agentic robot systems, since deep reasoning via 3D programs sits in the pipeline and is not obviously distillable to VLA speed (a VLA is a model that turns camera images and text into robot actions). Separately, Max Zhaoshuo Li said Astra cracked RoboLab with a near-perfect score, and Ken Goldberg pushed back, noting the authors have no previous publications and asking why the work is anonymous. Omar Rayyan also reported Astra outperforming all open-source VLA and world-model baselines on a subset of the zero-shot MolmoSpaces-v1 benchmark.
Gave GPT-6 Astra the task to plug in an ethernet cable. It did it after 2 hours and 4 human prompt interventions.
This is potentially major news but the authors have no previous publications…Why is it anonymous?
How much does it cost for this single run?
Also on the timeline
ETH Zurich's autonomous excavator stacked a 65-meter stone wall
Lukas Ziegler resurfaced ETH Zurich's Heap excavator, which scanned and placed over 900 stones, some above 1,000 kilograms, into a 65-meter dry-stone wall six meters high. The project ran from 2016 to 2025, so this is old work getting a new audience, not a fresh result.
A 192-neuron worm connectome drives a toy robot duck
Jiafei Duan trained a Microduck controller using a 192-neuron subset of the C. elegans male CNS connectome, with the modeled neural activity producing exploratory head and body movement. He calls it fun stuff, not a result, and no performance numbers were reported.
Palmer Luckey says Scoble charges $2,000 a post; Paxton denies doing it
Palmer Luckey said Robert Scoble DM'd him pitching Anduril on paid posts at $2,000 each, more for videos and blogs. Chris Paxton replied that when he shills for startups it is because he thinks they are cool, which is a useful thing to know given how much of this week's coverage rides on friendly quote-tweets.