I decided to review and explicitly post about the limitations of all my previous papers because I believe this is the fundamental driving force behind research, especially in this era of PRs and bubbles.
3.7 Flash brings a big jump in agentic performance and coding accuracy. To demonstrate, we set up a 3-agent team to autonomously train a robotics control model from scratch.
We hope you like 3.7 Flash, and you can read more here: blog.google/innovation-and…
🎯 Limitations:
In a setup with sufficient real data, the benefit of using sim as model prior is limited.
MBRL on real robots either has simplified dynamics (kinematics in this paper), or limited performance gain (Robotic World Model) due to the max domain gap it can handle.
🚀Introducing our new work: Learning More With Less: Sample-Efficient Model-Based RL for Loco-Manipulation
➡️sites.google.com/view/learning-…
🤖We show how model-based RL can learn complex robot dynamics efficiently — even on black-box hardware like Boston Dynamics Spot.
This is absolutely the best experience for me too @GoogleDeepMind!
People are extremely approachable and open to talk, even if I am just interning. The transparency built by @Google can be rare to see in even academic labs.
I thought the 1:1 with Jeff Dean was going to be the interesting part.
I was wrong.
What happened the next day is something I’ll probably remember for the rest of my career.
I posted the photo in my Google team chat and everyone in the office lost it.
"Wait, you met Jeff