"I don't want my robot triple-jumping in real life." Like in Halo
My conversation with Spencer Huang @NVIDIARobotics : the move from VLA → world action models, why game data breaks on physics, NVIDIA's sim-first stack, and why agents — not full autonomy — ship real robots
Cosmos 3 just released: from single Starbucks photo to full robot behaviour video!
From single photo I simulated 10 different behaviours just to what’s possible with robot in Starbucks environment!
@nvidiaomniverse@NVIDIARobotics
This scene has ground truth! The ground is created by https://t.co/jiZpxBuoEa! We use it to compare how well Cosmos generation reasoning is from single image to video! I have analysed it for generation, reasoning, manufacturing predictive maintenance!
Simulating different robot use case with NVIDIA Cosmos — all inspired by a single Starbucks scene 🤖
One image in. Grasping, serving, pouring, even a little "hello" wave out. World models are getting wild.
Every clip generated from one render. 👇 @NVIDIARobotics@nvidia
Automated real2sim pipeline is finally real
by @PalatialSim
I tested it with Isaac Sim and Franka Panda Arm— placing a fork into a dishwasher.
Both assets? Generated from scratch by Palatial in minutes. @StevenRSim@PalatialSim@nvidiaomniverse
Here's the full workflow: