I really like this; In my opinion I have strong conviction the industry is moving in this direction. Truly safe and capable robots will have to learn policy from thousands to millions of diverse Human Demonstrations - great work Shivam Vats @ CoRL2025

Fingers crossed this is true, looking forward to seeing what you built Bernt Bornich

everything that can be done on a computer will be done by a computer they are not operating excavators, they are training a model for autonomous excavators

A glimpse of a very exciting future ahead of us

Really love these demos, Agility is making some really strong progress on RL

Indeed

Incredible things are being built, take note garlicrobot.com

Looking fwd to presenting this talk Google next Thurs at noon. It will be live in person in Mountain View CA (not online) but is free and open to the public: How to Close the 100,000 Year “Data Gap” in Robotics rsvp.withgoogle.com/events/how-to-…

The state of affairs for deployable robotics today is significantly worse than much of the AI community who doesn’t work with Physical AI believes it to be

IKEA makes pretty good portable workstations

Incredible, I wonder what impact this will have on the open source robotics landscape

🍌

It seems like it still struggles with spatial temporal consistency here, i think we are still a couple of generations off robust world models but the progress is stunning. I imagine these will be extremely useful for bootstrapping generalist robot action sets.

Let’s go!

imo a problem in early-stage robotics: > Awesome demos on X > Yet, in reality, there is a big deployment gap (i.e. not as autonomous + scalable as in demo) > Very few can talk about the deployment gap (undermines reputation) > deployment gap stays underserved, harder to scale

If you have reservations, don’t - this isn’t how post-training is being done today but it will be in the future

.Epoch AI tracked every major robotics model (params, compute, training dataset size, simulation sizes, etc.) and found the largest models using barely ~1% of frontier AI compute, held back mostly by data scarcity. Link to data repo below. 👇🏼I found it very useful to run

A bullish signal for humanoid robotics is that VLAs are ~working. Different labs are showing that the same model can effectively generalize to various tasks. The conclusion is that we have an architecture on our hands that will take us a long way; we just need OOMs' more data.

🔁This is the most impressive transfer result I've seen: raw images to raw actions, across robots with different cameras and action spaces and ... We use a novel mechanism called Motion Transfer to learn across pre-training embodiments: no explicit alignment required!

Guterres really did drop a very gentle mogging here