🔥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

User-aligned robot representations are best learned from user data, but how do we collect this data without onerous labeling processes? We can learn from data that users are ✨intrinsically motivated✨ to produce through exploratory search 🔎! 🧵1/6

We see so many startups offering robot data collection -- how come we don't see ones focussing on real robot evals as a service? There should be real-robot benchmarks for all organisations to test their VLAs on.

I recently wrote a post about MILE for USC Robotics blog — check it out here: rasc.usc.edu/blog/mile-mode… Feel free to reach out if you have any questions or thoughts! See you at ICRA 🤖🙂

Arrived at #ICRA2025 and I'll be presenting my ReMEmbR work with NVIDIA on Tuesday! Happy to chat with people on robot memory, evaluation, language+robots, and reward learning (more to come soon on this one 😉)!

Reward models that help real robots learn new tasks—no new demos needed! ReWiND uses language-guided rewards to train bimanual arms on OOD tasks in 1 hour! Offline-to-online, lang-conditioned, visual RL on action-chunked transformers. 🧵

Jesse is such a naturally talented advisor that I didn't demonstrate anything about advising, just gave some language instructions. Do you know who else is so good? The model. Check it out 👇🏻

By the way, Jesse graduated last week and will move to Allen School for a postdoc. Be on the lookout for him when he enters the job market in a few years 🌟

Allen School Jesse Zhang Another shoutout goes to Jiahui Zhang , Yusen Luo and Abrar Anwar , who did most of the work. Yusen will apply for PhD in the next cycle, so another great candidate to be aware of. Abrar and I are both at #ICRA2025. We will be happy to chat with anyone about the work :)

Very very exciting to have Jesse Zhang join us at UW soon! He's done some incredible work - I'd recommend reading rewind-reward.github.io! Congratulations on a fantastic Ph.D. Jesse Zhang 🎉

Reward learning (like DYNA) has enabled e2e policies to reach 99% SR but (1) generalization to new tasks and (2) sample efficiency are still hard! ReWiND produces better rewards for OOD tasks than SOTA like GVL & LIV from Jason Ma that inspired us! 🌐: rewind-reward.github.io

Excited to share our work on continual, flexible, active, and safe robot personalization w/ Tom Silver, Ziang Liu, Ben Dodson & Tapomayukh "Tapo" Bhattacharjee. Also: Tom Silver is starting a lab at Princeton!! I HIGHLY recommend joining — thoughtful, kind, and an absolute joy to work with!

Wondering how to get more from your robot finetuning datasets? MILE extracts more out of intervention-based demos during training, giving you more bang for your buck per demonstration. Read more in the new USC Robotics blogpost! rasc.usc.edu/blog/mile-mode… Yiğit Korkmaz Erdem Bıyık

rewind-reward.github.io ReWiND: Language-Guided Rewards Teach Robot Policies without New Demonstrations LLM generated instructions z + demo -> learning-based reward model (progress) R(o,z) -> optimize policy via RL online

How can non-experts quickly teach robots a variety of tasks? Introducing HAND ✋, a simple, time-efficient method of training robots! Using just a **single hand demo**, HAND learns manipulation tasks in under **4 minutes**! 🧵

**When AIs Start Rewriting Themselves** Darwin G​​ödel Machine: Open-Ended Evolution of Self-Improving Agents The Darwin G​​ödel Machine can: 1. Read and modify its own code 2. Evaluate if the change improves performance 3. Open-endedly explore the solution space 🧵👇

VLAs have the potential to generalize over scenes and tasks, but require a ton of data to learn robust policies. We introduce OG-VLA, a novel architecture and learning framework that combines the generalization strengths of VLAs with the robustness of 3D-aware policies. 🧵

Cool to see more real-world RL! Great job!