Mengdi Xu
@mengdixu_
Followers
2K
Following
914
Media
6
Statuses
80
Assistant Prof. @Tsinghua_Uni. Postdoc @StanfordSVL. Ph.D. @CarnegieMellon. Prev. @GoogleDeepMind. Learning and Robotics.
Joined November 2017
How can we scale robot data to cover diverse household scenarios? ๐ก๐ณOne promising direction is generating large-scale bimanual mobile manipulation data in simulation! Excited to introduce MoMaGen, which is a scalable pipeline that automatically generates diverse long-horizon
We are excited to release MoMaGen, a data generation method for multi-step bimanual mobile manipulation. MoMaGen turns 1 human-teleoped robot trajectory into 1000s of generated trajectories automatically.๐ Website: https://t.co/DYKvqY4bII arXiv: https://t.co/lDffi0FXHl
2
8
81
AIโs next frontier is Spatial Intelligence, a technology that will turn seeing into reasoning, perception into action, and imagination into creation. But what is it? Why does it matter? How do we build it? And how can we use it? Today, I want to share with you my thoughts on
168
626
3K
I will join Northwestern University Computer Science as an Assistant Professor in Fall 2026! I am actively recruiting PhD students and seeking collaborations in robotics, human-robot interaction, brain-computer interfaces, cognitive science, societal impact of AI & automation,
67
207
1K
Thanks to everyoneโs interest in BEHAVIOR so far! We have received several questions, and I am trying to answer some of them here: 1. ๐How are tasks defined in BEHAVIOR? BEHAVIOR tasks are written in BDDL (BEHAVIOR Domain Definition Language). Unlike geometric, image/video, or
behavior-robot-suite.github.io
Streamlining Real-World Whole-Body Manipulation for Everyday Household Activities
(1/N) How close are we to enabling robots to solve the long-horizon, complex tasks that matter in everyday life? ๐จ We are thrilled to invite you to join the 1st BEHAVIOR Challenge @NeurIPS 2025, submission deadline: 11/15. ๐ Prizes: ๐ฅ $1,000 ๐ฅ $500 ๐ฅ $300
3
14
103
Very excited to announce the 1st BEHAVIOR Challenge! ๐๐ค With so many recent advances, how well do todayโs general-purpose models handle household tasks that are both practical and technically challenging? We invite you to submit your work and help benchmark the field. Can your
(1/N) How close are we to enabling robots to solve the long-horizon, complex tasks that matter in everyday life? ๐จ We are thrilled to invite you to join the 1st BEHAVIOR Challenge @NeurIPS 2025, submission deadline: 11/15. ๐ Prizes: ๐ฅ $1,000 ๐ฅ $500 ๐ฅ $300
1
7
63
Excited to share that ROSETTA won the Best Paper Award at the CRLH workshop @RSS! ๐ Huge kudos to the team, and many thanks to the organizers for a fantastic workshop! Iโll also be at the HitLRL workshop today. Happy to chat about building robots that better understand, assist,
0
0
11
ROSETTA won best paper at CRLH @ RSS and will be at HitLRL tomorrow! Thanks to @James_KKW, Jerry Chan, Roger Dai, @ManlingLi_, @mengdixu_, @RuohanZhang76, @jiajunwu_cs, and @drfeifei! Website: https://t.co/Sel0u8XoUB Code: https://t.co/A45CnR1HAK Paper:
0
5
21
Iโve always been thinking about how to make robots naturally co-exist with humans. The first step is having robots understand our unconstrained, dynamic preferences and follow them. ๐ค We proposed ROSETTA, which translates free-form language instructions into reward functions to
๐ค Household robots are becoming physically viable. But interacting with people in the home requires handling unseen, unconstrained, dynamic preferences, not just a complex physical domain. We introduce ROSETTA: a method to generate reward for such preferences cheaply. ๐งตโฌ๏ธ
4
5
52
How to scale visual affordance learning that is fine-grained, task-conditioned, works in-the-wild, in dynamic envs? Introducing Unsupervised Affordance Distillation (UAD): distills affordances from off-the-shelf foundation models, *all without manual labels*. Very excited this
9
112
437
๐คIntroducing TWIST: Teleoperated Whole-Body Imitation System. We develop a humanoid teleoperation system to enable coordinated, versatile, whole-body movements, using a single neural network. This is our first step toward general-purpose robots. ๐ https://t.co/ScrdX8ImNF
16
94
438
๐ Milestone Release! AReaL-boba, our latest #RL system! https://t.co/xmZe676YIZ
#AI โข data/code/model ALL๐ฅ #OPENSOURCE โข Full #SGLang & 1.5x faster on 7B RL โข SOTA 7B math reasoning: 61.9 AIME24 & 48.3 AIME25 โข 200-sample 32B tuning match QwQ on AIME24 @Alibaba_Qwen 1/3 ๐
7
42
113
Check out Yunfanโs work on solving real household tasks! Very cool to see the robot not only grasping but also skillfully using different parts of its body to manipulate objects.๐ช๐ค๐
๐ค Ever wondered what robots need to truly help humans around the house? ๐ก Introducing ๐๐๐๐๐ฉ๐๐ข๐ฅ ๐ฅ๐ผ๐ฏ๐ผ๐ ๐ฆ๐๐ถ๐๐ฒ (๐๐ฅ๐ฆ)โa comprehensive framework for mastering mobile whole-body manipulation across diverse household tasks! ๐งน๐ซง From taking out the trash to
2
1
31
๐Two weeks ago, we hosted a welcome party for the newest member of our Stanford Vision and Learning Labโa new robot! ๐คโจWatch as @drfeifei interacts with it in this fun video. Exciting release coming soon. Stay tuned! ๐๐
9
27
212
Aligning object physical properties in sim with real is crucial to close sim2real gap, especially in nonprehensile manipulation. We propose CAPTURE which adapts simulator w/o gradient updates by treating real and sim rollouts as contexts. Please check out Xilunโs ๐งตfor details!
๐ค What if robots could adapt from simulation to reality on the fly, mastering tasks like scooping objects and playing table air hockey? ๐ฅ๐ Iโm thrilled to share that our work, "Dynamics as Prompts: In-Context Learning for Sim-to-Real System Identification," has been accepted
1
2
40
The ultimate test of any physics simulator is its ability to deliver real-world results. With MuJoCo Playground, weโve combined the very best: MuJoCoโs rich and thriving ecosystem, massively parallel GPU-accelerated simulation, and real-world results across a diverse range of
37
186
909
Excited to see that Genesis is officially released. Congratulations to the team! Looking forward to seeing how Genesis will advance robot policy learning, data generation, policy evaluation, and more!
Everything you love about generative models โ now powered by real physics! Announcing the Genesis project โ after a 24-month large-scale research collaboration involving over 20 research labs โ a generative physics engine able to generate 4D dynamical worlds powered by a physics
1
0
39
Stay tuned!
[1/4] ๐Sneak Peek: SPARK in Action! ๐ฆพ Previewing Safe Protective & Assistive Robot Kit (SPARK)โa modular toolbox designed to enhance safety in humanoid autonomy and teleoperation. Safety isn't just a featureโit's the foundation for humanoids to truly integrate into human life.
0
2
13
Very excited to share with you what our teamย @theworldlabs has been up to! No matter how one theorizes the idea, it's hard to use words to describe the experience of interacting with 3D scenes generated by a photo or a sentence. Hope you enjoy this blog! ๐คฉโค๏ธโ๐ฅ
Weโve been busy building an AI system to generate 3D worlds from a single image. Check out some early results on our site, where you can interact with our scenes directly in the browser! https://t.co/ASD6ZHMwxI 1/n
78
264
2K
I'm building a new research lab @Cambridge_Eng focusing on 4D computer vision and generative models. Interested in joining us as a PhD student? Apply to the Engineering program by Dec 3 ๐๏ธ https://t.co/SDJEz2XiZp ChatGPT's "portrait of my current life"๐ https://t.co/qcnSgqYMWr
4
43
218