🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

What about mobility, you ask? One of the few places we've deployed reinforcement learning is in multi-agent pathfinding (MAPF), where we learn heuristics to efficiently plan for thousands of robots, minimizing congestion. amazon.science/latest-news/ho…

We're also very proud of our work in semantic scene understanding, led by the Canvas team in Boulder, which enables us to build semantically-aware 3D scenes and costmaps we use to navigate legibly and safely in a crowd. amazon.science/latest-news/th… (9/n)

This is my favorite program Amazon Science and I’m so proud to be working with a stellar team to tackle it. Picking and stowing in dense clutter quickly but gently is the next big manipulation challenge.

First post pandemic concert, featuring The Flaming Lips, robots, and a lot of confetti!

I wonder if the reason why we are excited by #ChatGPT’s plausible-sounding nonsense is because a lot of research and discussion nowadays is just that.

High density stowing with robots is one of the big open robotic manipulation we've been addressing Amazon Science. Read more!

After a wonderful 4 years Amazon building and scaling Robotics and AI for Fulfillment, I’m super excited to be starting a new journey with Kyle Vogt and @Cruise inventing algorithms, and partnering with Operations to scale autonomous mobility! Looking forward to this new adventure!

This is a main reason I've been excited about working @Cruise with Kyle Vogt - there's no substitute to driverless miles driven on real roads if you want to really learn to build and scale AVs worldwide.

An excellent example of Diffusion in action is when AI influences repost a repost a repost … of your paper until what’s remaining is pure white noise.

If you’re interested in getting Reinforcement Learning to work on real robots on complex manipulation tasks that even human experts fail on, and have an ok sim and a behavior policy, read our system paper for all the tricks to learn a superhuman policy.

Congratulations to a wonderful team on a project that’s very dear to my heart: empowering those in need of care via assistive robots. Looking forward to more!

That's some super impressive tech, but it's certainly not a driverless car.

Excited to be visiting UCI Engineering for the first time next week!

I am all in! Let’s go!

Super impressed by the physical dexterity demonstrated by all of these robots. But I’d add to Rodney Brooks’s comment that “imitation learning” without generalization is merely action replay, akin to what one can achieve with a teach pendant.

I too offer my personal services as a grumpy old professor to accuse your enemy with “this is never going to work on a REAL robot”.

The best way to resolve the beef between Elon Musk and Yann LeCun is to use a LLM. Can’t think of a better use of AGI than that.

Most interesting tasks are data scarce. We've spent the past year thinking carefully about leveraging structure, here Lipschitz continuity of the underlying unknown dynamics, to generate synthetic labels that enable superhuman fine manipulation. Read on.

Our work on representing traffic flow via topological braids is featured on the cover of the International Journal of Robotics Research! Read the paper here: journals.sagepub.com/doi/full/10.11… Joint work with Christoforos Mavrogiannis and Jonathan DeCastro.

I’ve been recently thinking much about data-efficiency in learning. While there’s no way to overcome the quadratic increase in mistakes with noninteractive IL like BC, a good representation that’s both attentive and generalizable significantly improves the constant.