Archive, page 2
- Flora goes open, Gecko's Navy deal hits $300M
A data engine chewing through 200,000 hours of video a day on 500 GPUs just had its weights posted for free. Plus the Navy quadrupled a robot inspection contract and Eka came out with a force-first manipulation model.
- The cornhole run, throw by throw
Google said its Apollo humanoid sank seven straight cornhole shots. The engineer who ran it then posted the whole log: throw #1 landed 85 cm short because the grip was slipping, and the fixes came one throw at a time.
- Adcock calls out a Hark refund story as a lie
Eric Lu said he emailed Hark support claiming the product failed, got a $30 pizza credit, and never downloaded the app. Brett Adcock replied with two words and a link from his team.
- A robot learned dexterous hands from minutes of play
Ruoshi Liu says minutes of human play, with no task demonstrations and no sim2real, were enough to train a dexterous manipulation policy. Jitendra Malik picked it up as his answer to what world models should actually be trained on.
- ASPIRE open-sourced after a model swap hit 99%
Jim Fan says his lab swapped the backend model in its robotics framework and saturated every benchmark to about 99% overnight. Elsewhere, the first reply to Tesla's newest clip was a researcher asking whether the footage is AI-generated.
- A G1 wrote its own code to fetch a Pringles can
An-Chieh Cheng says her team's G1 humanoid found the pantry and handed over a Pringles can with no pre-built map and no training for the task, because a coding agent wrote the program as it went. Separately, the SONIC controller moved onto full-size humanoids.
- Gemini 4 Argon sank seven cornhole shots on Apollo
Google says it showed its model a video of a human playing cornhole, and the Apptronik Apollo humanoid then hit seven straight. Takahiro Miki, who ran the demo, posted the blow by blow, including a throw that fell 85 cm short because the grip slipped.
- Gemini 4 Argon is driving humanoids, Google says
A world model trained on 30,000 hours of first-person video learned exactly where hands go and still could not say what happens to the things they touch. Going to 30,000 hours did not close the gap, and the field spent the day arguing about why.
- Zero-shot dexterity, and 300+ Cybercabs in Austin
A robotics startup says it went from zero to $100M in annual recurring revenue in six months, and nobody outside has checked a single number. Meanwhile two labs taught robots to thread nuts and catch boxes with no real-robot training data at all.
- Did those NEO orders ever actually ship?
Keerthana Gopalakrishnan answered a friendly post about bringing NEO to India with one question: did you ship last year's orders? Quiet Sunday otherwise, plus a small fight over whether controllers still matter.
- Tesla: 300+ Cybercabs in Austin, 8x in 30 days
A University of Washington group says its robots mesh gears and thread nuts with zero real-world training data, all clips at 1x. Meanwhile Tesla says Austin now has more than 300 Cybercabs, eight times as many as a month ago.
- Hark hits 281K daily users, and a humanoid catches boxes
Brett Adcock says Hark pulled 281,165 daily active users and 4.3 million messages in its first 48 hours, enough traffic to knock it over. Meanwhile a Unitree G1 learned to step sideways and catch boxes thrown at it.
- Adcock: Figure lifts way more than 50 lbs
Brett Adcock told Chris Paxton his humanoid payload chart was wrong, then fixed his own typo to say Figure can carry much heavier than 50 lbs with full loco-manipulation. Paxton's reply: the source ChatGPT pulled was wrong.
- Workhorse: a G1 humanoid learns from humans, no teleop
A Unitree G1 topples and climbs a suitcase, catches a thrown box and kicks boxes into place, learned from human demos with no teleoperation. Meanwhile Chris Paxton says neither Figure nor Optimus can lift what a warehouse worker lifts every day.
- Adcock ships Hark Pro with a 100,000 free plan cap
Brett Adcock's Hark Pro is live on web, iOS and Android, with the first 100,000 sign-ups getting the paid plan free. Also: a new humanoid built to walk into oil and gas disasters so people do not have to.
- Dyna's uncut laundry hour hits a dryer-timer question
Dyna Robotics called its hour-long laundry video uncut, and then Steve Crowe asked why the dryer counts down from 32 minutes to 1 in about 30 seconds of real time. That one reply is the week in robotics: the demos got longer, and so did the questions.
- Adcock: Figure's first hand was the wrong choice
Brett Adcock says the tendon-driven hand Figure started with "ended up becoming the wrong engineering choice," and he is now all in on the high-joint-count F.04 hands. Plus a tactile sensor you can print for five dollars.
- Digit's IROS demo ran on 500 teleop demos
The Agility Digit demo everyone liked at IROS turns out to rest on about 500 real teleoperated demonstrations and a sim-trained controller, hitting roughly 80 to 90 percent in a booth full of crowds and bad lighting. The person who built it says it is still research, not production.
- Digit ran mobile manipulation for 12 hours straight
Chris Paxton says the Agility Digit demo he ran at IROS went about 12 hours, autonomous, in a place that was not in its training data. Elsewhere, Yifan Hou says most of Atlas's new hand motions could not be teleoperated even if you tried.
- Boston Dynamics gives Atlas hands that lift a fridge
Atlas got 13-degree-of-freedom hands that are directly actuated, not tendon driven, and Chris Paxton says they can still carry a 100+ lb fridge. His read on the tendon-driven competition: it may be nerd-sniping.
- Figure flew its old robots to Finland to melt them
Brett Adcock says Figure trained its F.02 humanoids to jump autonomously, shipped them to Finland, and had them leap into a 75-ton electric arc furnace. Arnold Schwarzenegger's reply: "Hasta la vista, F.02."
- Dyna's new robot does an hour of laundry, uncut
Dyna Robotics says its new semi-humanoid Taku ran an entire hour-long laundry room workflow in one uncut take. Meanwhile a new public benchmark put four robot models through 14,400 real-world rollouts.
- Eidon AI shuts down: robot data has no buyers
Ted Xiao says the robotics data market has a single digit number of real buyers, and one company just closed because of it. Meanwhile Jiafei Duan posted unedited real-time footage of four frontier models running robot arms, and the field's verdict was: slow.
- Astra in real time is slow, Jiafei Duan says
Jiafei Duan posted an unedited rollout of GPT-6 Astra driving a robot and told viewers he applauds anyone who makes it to the end. Latency, not capability, is his argument against deployment.