Hongsuk Benjamin Choi

Humanoid robots learn from humans 🙌

Full episode dropping soon! Geeking out with @arthurallshire @redstone_hong on VideoMimic videomimic.net Co-hosted by @chris_j_paxton @micoolcho

🕹️

Five Golden Nuggets from this talk: 1. Pretrained finetuned vs single task policy: - Because of training on different tasks the pretrained policy has more recovery behaviours. - Somehow the visual-action mapping across different tasks and environments leads to this behaviour.…

Sharing a few more demos of our first #LargeBehaviorModel (LBM) at TRI. 1/🍎This model enables a robot to learned to core and cut an apple into multiple slices autonomously. We trained our diffusion-based LBM on almost 1,700 hours of robot data, conducted 1,800 real-world…

Congrats, OpenAI Team! And yes, you win our bet, Sanjeev. FYI for the curious: I had bet against "Some AI model will be able to perform at IMO gold level by May 1 2026".

For everyone interested in precise 📷camera control 📷 in transformers [e.g., video / world model etc] Stop settling for Plücker raymaps -- use camera-aware relative PE in your attention layers, like RoPE (for LLMs) but for cameras! Paper & code: liruilong.cn/prope/

Flow Q-learning (FQL) is a simple method to train/fine-tune an expressive flow policy with RL. Come visit our poster at 4:30p-7p this Wed (evening session, 2nd day)!

Ep#20 with @arthurallshire @redstone_hong on VideoMimic videomimic.net Co-hosted by @chris_j_paxton @micoolcho

This is a great compliment! Our real-to-sim code is now available. It can recover both the environment and target motion from videos. github.com/hongsukchoi/Vi…

Action chunking is a great idea in robotics: by getting a model to produce a short sequence of actions, it _just works better_ for some mysterious reason. Now it turns out this can help in RL too, and it's a bit clearer why: action chunks help explore and help with backups. 🧵👇

If you're interested in working on learning from simulation, direct from perception and training at scale for multi-fingered hands please reach out to me and my colleagues. We are looking for talented researchers who have worked or have experience working on those problems.

🚀Thrilled to share what we’ve been building at TRI over the past several months: our first Large Behavior Models (LBMs) are here! I’m proud to have been a core contributor to the multi-task policy learning and post-training efforts. At TRI, we’ve been researching how LBMs can…

We build Cosmos-Predict2 as a world foundation model for Physical AI builders — fully open and adaptable. Post-train it for specialized tasks or different output types. Available in multiple sizes, resolutions, and frame rates. 📷 Watch the repo walkthrough…

ummm… As a robotics PhD student, I’m genuinely worried that the problem I find important now will be solved in the next 2 years—by MORE DATA, without any need to understand the underlying structure. And this happens in many areas😂

Amazing @taeinkwon1

#CVPR2025

Q-learning is not yet scalable seohong.me/blog/q-learnin… I wrote a blog post about my thoughts on scalable RL algorithms. To be clear, I'm still highly optimistic about off-policy RL and Q-learning! I just think we haven't found the right solution yet (the post discusses why).

1/ How should RL agents prepare to solve new tasks? While prior methods often learn a model that predicts the immediate next observation, we build a model that predicts many steps into the future, conditioning on different user intentions: chongyi-zheng.github.io/infom.

Cudos to VideoMimic team:)