We found a way to scale humanoid control that reduces GPU-hours by more than 1000×.
Across 129,514 motions, SuperMimic surpasses the state of the art (128 GPUs, 7 days) with 1 GPU in 3 hours.
Project: supermimic.github.io
Demo: supermimic.github.io/demo/
This 😈 E-Atlas fan gave his G1 a makeover and bullied it.
Please believe that I am loyal to you, and punish him (should the God of AI ever read this in the future). Lol.
Meet Workhorse 🐴: we study how to learn whole-body loco-manipulation skills from human demos without teleop.
Here is my favourite video: we tried everything to make it fail🏒.
One take, real-time, fully autonomous.
Absolutely you can. Game rules were changed since AI tools alleviate the gap between undergrads/MS and PhD in terms of knowledge and coding. The only advantage of PhD (in my major, Robotics, again) is the experience on hardware. And maybe better research taste, built through long-term research. But many undergrads now start doing research as sophomores and already have a solid track record by the time they apply to PhD programs. Just look at PhD applications from the past two years. It's insane. They're even stronger than the CVs of PhDs who graduated 5 to 10 years ago.
Read this and felt nothing. In my field (robotics), I don't think academia can match industry in any aspect of research, especially when your work (1) needs compute; (2) needs data; (3) needs robot hardware; (4) can be massively accelerated by LLMs
Btw I'd really love to know which PhD advisors reimburse their students for the LLM costs they pay out of pocket to push the lab's research forward. And how many tokens do they actually have to work with?
Of course, industry and academia were never enemies. One of the few perks academia has left is that you can sometimes put years of funding, guilt-free, into ideas that have zero chance of shipping soon. You're free to ignore the mainstream and explore new directions. With AI/robotics this saturated, I think that kind of exploration is worth far more than using LLMs to take convex combinations of existing knowledge and churn out toilet paper. It could well lead to a Transformer-level breakthrough and push AI's ceiling higher. The problem is that getting limited funding to the right researchers is hard, and academic freedom is scarce even in academia. In the end, it's purely a fight for ideals, for something metaphysical.
Students sometimes ask me if it still makes sense, in this accelerating age, to pursue a PhD in AI.
Perhaps counterintuitively, I think it's a great time to do so.
I wrote up some thoughts on this here: web.mit.edu/phillipi/www/w…
More importantly, the policy was running without the motion capture. Unlike recent other demos you can see on X, it's fully autonomous! We would post the preprint and code soon.
Dribbling ⚽️ is much harder than kicking. A kick is one touch. In dribbling, every touch decides where the ball rolls and whether the robot can still reach it. Small heading errors also add up from touch to touch.
Our G1 weaves through poles at 1 m/s and dribbles straight at over 1.2 m/s, keeping the ball close the whole time. One policy, human-like motion, zero-shot from sim.
So proud of this work with @ZihanChen_77@zitengji@ChuyeHong and the team of HRG!
Fast dribbling. Tight turns. Repeated direction changes. Human-like motion. ⚽️
We present a Unitree G1 system for sustained humanoid dribbling beyond simple straight-line ball pushing.
The robot continuously controls the ball while running through slalom courses and large
The September 2026 issue of Science #Robotics is out!
This month's cover depicts AthenaZero, a robot with humanlike mass and dynamic capabilities that enable it to throw, catch, and hit a baseball quickly. Learn about this research and more: scim.ag/46WYd6J
“AI agents will outperform humans at almost all jobs by 2026–2027.” - The forecast is everywhere.
So we built the exam to test that claim, on real labor-market aligned work. On the hardest tier, top agents pass 2.6%.
Meet Agents' Last Exam (ALE), a rolling benchmark measuring whether agents can actually do real jobs. 🧵👇
That’s one of the reasons why I recently prefer RA-L more. If there is a double-anonymous journal that can finish one-round constructive rebuttal and give you feedback within 3 months, then Conferences like ICRA and IROS will be less attractive. Especially RA-L paper can also be presented at these conferences.
CoRL 2026 is cutting review cycles to <3 months to keep pace with the rapid progress in robotics. @ieee_ras_icra@IROS2025 — maybe it’s time to rethink too.
It makes little sense to discuss papers at conferences a year after submission, when they’re already outdated.
@TairanHe99@ZhiSu22@ieee_ras_icra@IROS2025 I personally think that would be hard to execute because they are from different publishers (IEEE and PMLR). Especially RAS is planning to have a new journal about robot learning.
Kindly remind that Zhi is applying his PhD and if you are looking for a student who is both talented and hardworking, please contact him!
He is the best undergrad researcher I have ever seen. He finished this work in only 3 months. As his roommate, he kept working til 3am everyday.
Amazing work by Zhi !
Researchers still don’t reach the ceiling of locomotion, humanoid robot can do more! Unfortunately I didn’t have it when I was a child. Jealous to the next generation :)
🏓🤖 Our humanoid robot can now rally over 100 consecutive shots against a human in real table tennis — fully autonomous, sub-second reaction, human-like strikes.
839 Followers 5K FollowingLLM Post-Training/ Recursive Self-Improvement for Reasoning
Co-founder/ MTS @ReasonCoreAI
PhD @UCLA, Bachelors @iitbombay
Led a team of ~500 PhD researchers
9K Followers 2K FollowingAssistant Professor @CMU_Robotics @SCSatCMU, leading the @LeCARLab. @Amazon Scholar @ FAR. I build generalist robots with agility. Ph.D. from @Caltech.
612K Followers 60K FollowingThe automated life | San Francisco/Silicon Valley AI, robotics, BCIs | Ex-Microsoft, Rackspace, Fast Company | Wrote eight books about the future.
7K Followers 73 FollowingProf @Stanford, Distinguished Research Scientist and AV research lead @nvidia. PhD from @MITAeroAstro. Robotics, autonomous systems, AI. Opinions are my own.
794 Followers 162 FollowingCS PhD at Stanford | Prev @physical_int, @ucsd_cse, @rai_inst, @StanfordSVL, @NUSingapore | Building intelligence across bodies and teams
5K Followers 568 FollowingAssistant Professor @Harvard SEAS @hseas, Lead the Harvard Computational Robotics Lab. #Robotics, #Optimization, #Control, #Vision, #Learning
1.1M Followers 212 FollowingOnly on X, don’t trust fake accs
AI/Semi Supply Chains Research
Nothing is investment advice. No paid promos; may trade/hold names disc, views my own.
601K Followers 3K FollowingNVIDIA Director of Robotics & Distinguished Scientist. Co-Lead of GEAR lab. Solving Physical AGI, one motor at a time. Stanford Ph.D. OpenAI's 1st intern.
12K Followers 580 FollowingCo-Founder of SceniX (now part of @theworldlabs) | Assistant Professor @Columbia @ColumbiaCompSci | Former postdoc @Stanford @StanfordSVL | PhD @MIT_CSAIL
6K Followers 1K FollowingAsst. Prof@ Tsinghua University, Co-founder @GalaxeaDynamics. Former Scientist@Waymo, MIT PhD’19. I research on multimodal learning, self-driving and robotics!
24K Followers 282 FollowingPioneering the future of robotics since 1979. We’re transforming industries and everyday life through cutting-edge innovation and world-class education.
6K Followers 199 FollowingSharpa is an AI robotics company dedicated to developing ultra-high performance robots and core components.
We Manufacture Time by Making Robots Useful.