Kaixuan Wang (minifull) retweeted
Any online optimization (IK, ID, MPC) for low-level control should be abandoned. A RL policy can fundamentally track more precisely, more robustly, and adaptively to unknown disturbances and uncharacterized noise, especially for low-cost hardware, even on a manipulator.
Kaixuan Wang (minifull) retweeted
Jev + Astra beats the Ender Dragon in Minecraft in 8 minutes 43 seconds! ⏱️
Cost less than $1 ($0.01 Jev, $0.96 Astra)
I open sourced the code and explain the harness setup below. This type of movement is only possible with Jev's near instant decisionmaking, and some continually learning skills from Astra.
Kaixuan Wang (minifull) retweeted
We put GPT-6 Astra in the RoboDojo. 🥋🤖
The RoboDojo Team conducted a comprehensive evaluation of GPT-6 Astra as an embodied agent, including:
• RoboDojo Sim & Real, compared with GPT-5.5 and DeepSeek-Flash
• Humanoid high-level control
• Dexterous piano playing with RoboPianist 🎹
• A systematic study of in-context learning (ICL)
Our key takeaway:
GPT-6 Astra demonstrates remarkably strong semantic and spatial understanding, together with impressive in-context adaptation.
At the same time, physical commonsense remains a clear bottleneck — revealing an important gap between understanding the world and truly reasoning about its physics.
Full report & demos:
robodojo-benchmark.com/repor…
@_wenbozhang (project lead), @wenhaocha1, @frankzydou, @JinWeiyang18434, @YutaoOuyang, @minifullcapsule, @x_h_ucb, @YutaoOuyang, @YueChen614
Kaixuan Wang (minifull) retweeted
We evaluated 30+ frontier embodied AI models.
The result is clear: current generalist robot policies are still far from robust real-world manipulation.
This is why we built RoboDojo.
Officially launching ManiTwin🤖 - We introduce an automated pipeline for annotated digital twin asset generation and present ManiTwin-100K, a manipulation-ready 3D object dataset with VLM-driven semantic annotations.
Project: manitwin.github.io/
Thrilled and happy to be advised/cooperating with Prof. Li! If you are passionate about robotics and embodied AI, definitely check this out.
Excited to share that I’ve recently joined the Chinese University of Hong Kong (CUHK) as an Assistant Professor in Mechanical and Automation Engineering! My research will continue to focus on embodied AI & humanoid robotics — legged locomotion, whole-body and dexterous manipulation, robot foundation models, and more.
Hong Kong offers a unique environment: next to mainland China’s rapidly growing robotics ecosystem, while also being an international and globally connected city. A great place to build impactful robotics research.
🎓Multiple PhD openings (legged, manipulation, aerial, dex hand, and more)
🗓️Deadline: Dec 1, HKT (www4.mae.cuhk.edu.hk/researc…)
🌍International applicants are encouraged to also apply for HKPFS (cerg1.ugc.edu.hk/hkpfs/index…)
Why the experiments are done with 1B, 6B and 7B models rather than 1B, 7B, 13B and even 70B models?
Is that cherry-picked?
Kaixuan Wang (minifull) retweeted
Introducing GEN-0, our latest 10B+ foundation model for robots
⏱️ built on Harmonic Reasoning, new architecture that can think & act seamlessly
📈 strong scaling laws: more pretraining & model size = better
🌍 unprecedented corpus of 270,000+ hrs of dexterous data
Read more 👇
Kaixuan Wang (minifull) retweeted
Imitation learning is great, but needs us to have (near) optimal data. We throw away most other data (failures, evaluation data, suboptimal data, undirected play data), even though this data can be really useful and way cheaper! In our new work - RISE, we show a simple way to *use all of this non-optimal data to robustify imitation learning* with minimal requirements beyond BC.
Key idea: use non-expert data to learn how to *recover* back to expert data with a minimal frills offline RL that works under sparse data coverage. Allows usage of *all* available data, not just expert data - never throw your data away!
Paper: arxiv.org/abs/2510.19495
Website: uwrobotlearning.github.io/RI…
A 🧵(1/10)
Congrats
Two weeks ago I passed my PhD thesis proposal 🎉 Huge thanks to my advisors @GuanyaShi & @ChangliuL, my committee, and everyone who has helped me along the way.
Last week I also gave a talk at UPenn GRASP on our 2-year journey in humanoid sim2real—reflections, lessons, and outlook. Great Q&A, now online: 🔗 youtube.com/watch?v=sRpx2hka…
During the visit, I had inspiring 1:1s with 10 professors and even an hour with GRASP Lab founder Ruzena Bajcsy. I asked her: after nearly 60 years as a robotics professor, what defines a truly great researcher? She said:
1️⃣ Don’t publish many papers
2️⃣ Don’t pick trivial problems
Two principles I’ll strive to follow for the rest of my PhD journey.
Kaixuan Wang (minifull) retweeted
Human data for scaling up mobile manipulation!! 🚶🏻👉🏻🤖🦿
Happy to be in the first batch of students
I am going to offer a new course DATA8014 Principles of Deep Representation Learning at HKU this Fall 2025 for incoming grads and senior undergrads. It will be based on a new textbook that we will soon release publicly. Hope that this will fundamentally change the teaching of deep learning and principles of intelligence in general. We are also planning to offer the same course at UC Berkeley in Spring 2026.
Kaixuan Wang (minifull) retweeted
Looking for the EXACT #CVPR2025 decision release time? Don’t ask, we don’t know either 😉 Good luck!
Timelines: cvpr.thecvf.com/Conferences/…
Cool
The ultimate test of any physics simulator is its ability to deliver real-world results.
With MuJoCo Playground, we’ve combined the very best: MuJoCo’s rich and thriving ecosystem, massively parallel GPU-accelerated simulation, and real-world results across a diverse range of robot platforms: quadrupeds, humanoids, dexterous hands, and arms.
Best of all? You can get started today with a single command: pip install playground
playground.mujoco.org/
Just glancing at the table of contents, I believe this book was definitely worth reading
