Figure lines up three humanoid generations
An AI agent built a working physics copy of a two-arm robot task from a phone video in 56 minutes, then fixed its own modelling error after the real robot disagreed. Plus Figure's hardware timeline in one shot and Eka's pick-rate numbers.
Read the whole week in one goOne GPU, three hours: SuperMimic's 1000x humanoid training
Shown this week
Figure puts three generations of its humanoid side by side
Brett Adcock posted three generations of Figure humanoids together in one shot. The post is a single line of text with the clip and no specs, no task and no performance claim attached.
AI agent builds a robot simulation from a phone video in 56 minutes
Lukas Ziegler flagged RoboSimStudio, where an AI agent built a physics twin (a simulated copy of the real setup) of a two-arm wiping task from a smartphone video, with no CAD and no calibrated cameras. He says Claude and GPT-6 Astra each built one independently and one finished in 56 minutes.
The real world found the modelling error and the fix went back into the twin.
Out in the world
Eka's pick numbers land: 1,500 an hour, under $1,000 a gripper
Numbers are now attached to Eka Robotics' $150M Series A, which Lukas Ziegler says closed at a $1.45B valuation led by a16z and F-Prime. He puts the vision-force-action system at roughly 1,500 objects picked per hour against roughly 600 for a human, with 95%+ reliable pickup. The custom gripper is listed under $1,000 and rated for 5M duty cycles.
force and touch close the last-millimeter problem vision alone cannot solve