OM-1 unplugs an Ethernet cable, and the video is 1x speed
Reward AI's new foundation model drew praise from Chris Paxton for a task most robots fumble.
The detail people latched onto in Reward AI's OM-1 launch was the playback speed. Lukas Ziegler watched the demo footage and pointed out that it runs at 1x, no time compression, which is unusual enough in robot demo videos to be worth saying out loud.
That matters because of what the robot is doing. Reward AI tested OM-1 on unplugging an Ethernet cable, which the company calls a deceptively hard task. Unlike a USB plug, an Ethernet connector locks in place and only comes out when the latch is pressed very precisely. The company says OM-1 solves it reliably.
Ethernet connectors lock firmly in place and release only when the latch is pressed very precisely
Chris Paxton, who has been quick to poke holes in other robot claims this month, was complimentary here. He called it a very difficult task for a robot and said the result is really good.
What OM-1 claims
OM-1 is Reward AI's first robot foundation model, announced as zero-shot generalizing to any robot, meaning it drives a new robot body without extra training for it. The company lists tabletop arms, industrial arms and humanoids. It says the model learned directly from human manipulation data with no teleop and no robot data, and claims close to human-level dexterity and efficiency plus multi-robot collaboration. We covered those training claims when the model first went up, earlier.
If the no-teleop part holds, it changes the economics of training. Today a new robot body usually means a new pile of teleoperated data, a human puppeting the machine for hours to give it examples. Learning from video of humans doing the task instead would skip that step for every body.
Shown versus claimed. What is on video is a robot pressing the latch and pulling an Ethernet cable out, at real speed, and one well-known researcher saying that is hard. The generalization across arms and humanoids, the human-level dexterity line and the no-robot-data training recipe are all Reward AI's own descriptions, with no outside test of them yet.
daaamn, just realised this is on 1x speed
This is really good; very difficult task for a robot