Excited to release an universal tokenizer that trains VLAs x5 faster! This is going to be a game changer for dexterous high frequency control like humanoids 🤖 This was a fantastic collab between Physical Intelligence and Berkeley AI Research led by rockstars Karl Pertsch and Kyle Stachowicz !

I guess that's what happens when you are so ahead of your time! Though I definitely don't miss using caffe to train my networks, specially having to write the forward & backward pass for custom layers in c++ 😅

RL and test-time compute have been hot topics lately. Curious about how to apply them to robot policies? We’ve released our training&eval code for V-GPS. I hope this can be a good starting point for those interested in test-time computation for robotics 😉 github.com/nakamotoo/V-GPS

One step closer to truly interactive multimodal robots 🙂 Congrats to all the robotics team at Google DeepMind for the cool work!

Thrilled to present my research on robot foundation models today in-person at ETH Zürich as part of the Robotics, Vision, and Controls Talks! Talk starts at 16:00 CET; ETH HG E 41, zoom link: robotics-talks.com Come talk to me if you are excited about robot foundation models!

The talk I gave last week at ETH Zurich Zurich🇨🇭 in-person is now online (thanks Robert Katzschmann for the invitation!). Check it out if you want to learn more about the ingredients to develop generalist robot policies 🤖 and our latest research at Berkeley AI Research! youtube.com/watch?v=z_TTw0…

Congrats to the Wayve team on the impressive release! Multi-cam world models with efficient video tokenizers 👌

Embodied chain of thought (ECoT) is a powerful tool to get VLAs to think through problems, but why does it work? In our new work, we analyze various lightweight ECoT-like strategies, including co-training, to see what is the "minimal" amount of reasoning that can boost VLAs 🧵👇

Congrats to Andy Zeng and Pete Florence on the impressive demos! 🎉 It's been an absolute pleasure working with Andy over the years, and I highly recommend applying to join their incredibly talented team!

Thrilled that our work on efficient robot tokenization was nominated for a Best Conference Paper Award at RSS 2025! I unfortunately won't have the chance of attending #RSS this year, but reach out to Karl Pertsch and Kyle Stachowicz if you want to learn more about FAST!