🔥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

just coded an RL env for Suika. try it out and let me know if your agent can get the watermelon! github.com/edwhu/suika_rl

So excited and grateful to share that I matched UC San Francisco for internal medicine residency today! Thank you to my mentors, family, friends, and fiancé Edward Hu for all of your support during my med school journey! Couldn’t have done it without you ❤️

bruh this RL shit is hard 😭

Pi0 really did work for us on the first try. No camera calibration, controller tuning, etc. The failure cases: missed grasps and risk-averse "hedging" behavior. Excited to see how the robotics community improves on this. At the very least, it will be a good baseline.

I'll make a tweet before ICLR'25, but this thread captures the essence well. Predicting the next token with transformers is great; but predicting 2 tokens = provably sufficient representation for planning. Let's do that and see what happens.

Sometimes, it's expensive to reset in RL (e.g. robotics). It turns out world models do pretty well here. Why? • learn reset policy in imagination for free • policy training in world model is resistant to distribution shift. Check out Zhao's TMLR accepted project!

Well deserved. Dv3 made my last RL paper super easy to tune for new tasks: just vary model size and UTD ratio. Downside is the complexity of code. In PPO, I need to tune clipping, entropy, lr, batch, per new task. But the code is simple. Use the right tool for the job!

Introducing Dynamism v1 (DYNA-1) by Dyna Robotics – the first robot foundation model built for round-the-clock, high-throughput dexterous autonomy. Here is a time-lapse video of our model autonomously folding 850+ napkins in a span of 24 hours with • 99.4% success rate — zero

⚠️ Attention: The site is currently down. Our engineering team is investigating. We will update as soon as possible. You can track progress here: status.overleaf.com Sorry for any inconvenience.

On my way to Atlanta to present ZeroMimic: Distilling Robotic Manipulation Skills from Web Videos at IEEE ICRA! Stay tuned for an in-depth post about how ZeroMimic distills zero-shot policies from web human videos. 🌐 Project site: zeromimic.github.io