Its been doing this for like two hours. RL whole body mobile manipulation, objects bought on Monday, never seen before

these consistency demos are way more impressive than cherry-picked videos.

It may look slow - but once you can tell it to do this 24/7 you can walk away, the data flywheel will give you all the real work data you need after a year/s of doing this Imagine 1-100-1,000 units running and + the faster they get the more data/hr you earn

@TheHumanoidHub Speed will come later. Just the fact that they can do this dynamically with novel objects is huge. Double the speed and improve the coordination, and I could see these deployed in grocery stores to stock shelves 24/7 or even to automate delivery pickup requests made online

how is perception handled? State estimation or is this zero shot rgb/depth policy?

Depth + open vocabulary object detector. Modular stack

walk could be better and it could be faster but it seems smooth and coordinated pretty cool

Yeah the walk could be faster. It's very stable and handles crazy arm motions while reaching and picking

Awesome! What kind of RL? Sim2real? Or real world RL? Learned from scratch or RL fine-tuning on top of existing policy?

Zero shot, sim to real!

Which exhibit hall is this in??

Hall 1, nviida booth