I've always been someone who follows my curiosity. This new funding gives us the opportunity to keep scaling our research, pursue even more ambitious ideas, and bring what we're building into the world. None of this would have been possible without the hard work of so many amazing people on our team.
On a more personal note, I’m deeply grateful to the friends and mentors who have supported and influenced me along the way. I’ve been very lucky to meet so many kind and wise people early in my career.
The future worth building is human. Onwards!
Genesis is entering its next phase.
Today we've raised a $500M Series A at $3.5B.
We can imagine far more than we can build physically today. Almost everything we are creating is still paid for in human time, by human hands. We want to remove the physical limits on human ambition.
We are building general-purpose robots with human level intelligence and capability. Ones that work reliably across entire workflows for years.
There are countless versions of possible futures where billions of robots are deployed everywhere, and we aspire to bring the brightest one that carries our hopes into reality.
Two things Xian has taught me over the past few months that I think are also illustrated really well in this interview:
1. Robot intelligence is a bounded local problem.
2. The real moat for a robotics company lies in its product, brand, and user experience.
Robots can do backflips. They still can't plug in a cable.
Everyday work is much harder than it looks. Gripping, pressing and handling things that bend and slip is most of physical work, and robots are still bad at it.
That is the problem @gs_ai_ is going after. Genesis is building general-purpose robots that can do this kind of hands-on work in the places people do it today, starting with industrial customers. We're thrilled to co-lead their Series A and to partner with @zhou_xian_ @theo_gervet and the team as they take it into the real world.
What drew us in is that Genesis is solving it full stack. The team builds the hardware, the data engine, the model and the simulator in house, because each one depends on the others. The hardware decides what data you can collect, the data decides how good the model can get, and the simulator is how you test all of it at scale. Getting any one of them right means getting them right together.
The team has rare depth across every one of those layers, and they have been focused on real deployments from day one.
I sat down with Xian and Theo to dig into all of this. The video is below, and our full thesis is in the comments.
Congrats to the whole Genesis team. Onwards.
1
13
872
Zu retweeted
Two years ago, we released Genesis, a universal physics engine for robotics.
Today, I’m excited to announce 🍎 Eden — a robot learning framework built from the ground up for agents, unleashing full potential of Genesis.
One task definition. From simulation to real robots. Sim-to-sim and sim-to-real just got easier.
Coding agents build tasks 1.6–4× faster than in Isaac Lab.
🧵 1/11
Tried this demo myself and asked it to pick up my phone. Also saw others trying it with very thin and deformable objects. Amazing generalizability and dexterity. Pure ego + tactile, no wrist cameras. Very first principles pilled.
Impressive! Looking forward to the full paper!
In behavior cloning, we mastered how to learn robotics from "robot-complete" data – but what happens to data that is not so perfect? Intelligence is often about learning from incomplete references; to come up with a "story" given disparate datapoints. Our Pick Anything demo shows an ability of this sort: the policy learned human-like dexterity from absolutely no human-robot interactions, no teleop or handheld tool data, just letting humans be humans and robots be robots. How did we do this?
Learning from humans without actions is often reduced to a semantics problem, a fancy way to prompt the robot. What, if anything, can robots learn about the dynamics from them? RL for manipulation has the opposite problem – often we are stuck with "technically correct" solution that are nevertheless aesthetically unsatisfying. In Pick Anything, we found a way to unite the two in a way for dexterous manipulation that I personally find quite beautiful (and I can't wait to share the paper with you!) I believe what we have here is the key to human level dexterity for robots, and all of that without needing any real robot data whatsoever.
For this Monday/Tuesday, you can visit our demo booth at CoRL (Griffin hall, demo 12) and ask me more questions! It's a labor of love from my past year of work, and I'm very excited to be finally able to talk about it.
Good ideas can come from anywhere. Opportunities should, too.
I've been thinking a lot lately about the relationship between academia and industry, and have had many thoughtful conversations with friends about how we can better support the next generation of researchers.
As more resources flow into robotics, I hope we can create more opportunities for ambitious people to pursue ambitious ideas. After all, the progress of a field shouldn't just be measured by the breakthroughs it produces, but by the opportunities it creates for the next generation.
Will be briefly at CoRL next week, hosting two events with @johnsonwang0810. Especially looking forward to meeting students and early-career researchers.
Khosla Ventures × Genesis Robotics Mixer:
luma.com/548fmnsq
CoRL After Party with General Catalyst, NVIDIA & friends:
luma.com/general-oxrj
Exciting times ahead for Genesis. More to share soon :)
Mahi is once again ahead of his time. Behavior cloning on real-world data has been the dominant paradigm, but simulation and reinforcement learning are poised for a renaissance.
Gradually, then suddenly.
I will be at CoRL, presenting Do as I Do, and a yet unreleased demo ✋🤖
My opinion on the future of robotics has shifted drastically over the past year. I believe we're headed towards an inflection point: soon, majority of our robot data will be created without robots (or even humans) in the loop. Why? Because scaling robotics at the speed of robots or humans is almost always going to be slower than scaling at the speed of compute.
We want robots that work in the real world. How do we stop making the real world our bottleneck? Looking forward to sharing my thoughts, and would love to hear yours if you are going to be in Austin next week.
16
1,490
Robots need to feel where they touch and how forces change. Introducing PanoTac: an open-source, dense 3D tactile sensor for <$10 USD.
Watch as the robot racks a test tube and pulls the stopper: PanoTac tells the tube orientation, and feels when the stopper comes free. 🧵
I won’t make it to CoRL this year, but our robot peeling work will be presented at Monday’s poster session! Paper, code and data are fully updated: toruowo.github.io/peel/
This work is almost a year old, and a lot has happened since. I checked this afternoon, and was relieved to find that peeling potato with a knife still isn’t something that Astra can do out of the box (yet)!
Anyways, most publications eventually become obsolete as new advances emerge, and AI is accelerating that cycle. While the specific methods used in this work might not last, I hope the idea does: a robot policy shouldn’t just get the task done, but also do it the way people prefer. How to incorporate human preferences into robot policies still feels underexplored, and I’m excited to see more work in this direction.
For any related discussion or follow-up, feel free to reach me by email :)
Peeling a potato is trivial -- until you try to make a robot do it with a knife.
This is actually one of the hardest problems in manipulation: contact-rich, force-sensitive, and success is subjective.
We taught a robot arm to peel with >90% success :D
toruowo.github.io/peel
Zu retweeted
I believe that we are entering an era of abundant robot data. Therefore, the story of finetuning large pretrained video models as robot world models is the story of the past. We should discover the optimal architecture and training recipe for scaling world model training on robot-native datasets. @_YimingDou and I have been validating this belief at Amazon FAR and have obtained some preliminary success. Proud to share this work with everyone!
Introducing DreamFAR, a world model trained from scratch using only robot data, without internet pretraining. It supports:
🎮 Real-time teleoperation
⏱️ Minutes of long-horizon rollout
🧩 Generalization across 1,000+ tasks
yimingdou.com/DreamFAR
🧵 (1/8)
Yes, contact is important and often overlooked. Simple workarounds that may suffice for parallel grippers don’t readily extend to dexterous hands. Nice to see this new paper tackling the problem with contact-anchored retargeting.
Also my favorite manipulation paper from FAR so far this year. Great work, @akashshrm02!
Thrilled to share CARD: Contact-Anchored Retargeting for Dexterous manipulation. So far, collecting robot data, teleop, UMI, kinesthetic teaching, has run into the same two problems: 1) the mapping is largely kinematics only, and 2) the human has to act through the robot and hope the interaction works out. CARD instead treats contact and forces as the interface, from recording to replay, which is what lets it handle dexterous tasks that motion-only methods struggle with.
🧵 below 1/
Zu retweeted
. @ToruO_O also is a great and authentic writer on other platforms and have great artistic taste too :)
“choosing and defining the right problems, and exercising judgment about what the results actually mean.
But then, that high-level taste and judgment don't just come from nowhere. In the past, it was precisely through the mundane trial and error -- the very process that is being automated away today -- that researchers developed the so-called "scientific taste".”
How should we approach research, when the very field we work on is going through rapid and potentially fundamental changes?
I recently gave a talk at the Stanford Vision Seminar, reflecting on my PhD journey and how AI is changing the way I think about research. I adapted part of it into a new blog post, organized around three questions:
1. What should we work on?
2. How should we work with AI?
3. What is a PhD for?
I hope some of my reflections are helpful. The talk ended with an open discussion, and I’d love to continue it here; curious how others are thinking about these questions.
Personally -- as much as there's FOMO, I also think life is very long, and we have time.
toruowo.github.io/blog/posts…
How should we approach research, when the very field we work on is going through rapid and potentially fundamental changes?
I recently gave a talk at the Stanford Vision Seminar, reflecting on my PhD journey and how AI is changing the way I think about research. I adapted part of it into a new blog post, organized around three questions:
1. What should we work on?
2. How should we work with AI?
3. What is a PhD for?
I hope some of my reflections are helpful. The talk ended with an open discussion, and I’d love to continue it here; curious how others are thinking about these questions.
Personally -- as much as there's FOMO, I also think life is very long, and we have time.
toruowo.github.io/blog/posts…
💜
Life update: I recently graduated from CMU RI, and last year I went through the academic faculty job market in robot learning. I was fortunate to receive a lot of help and support from many people throughout the process.
Compared to industry job applications, there are relatively few resources available online about the faculty application process. So I decided to open-source all of my application materials from last year here:
docs.google.com/document/d/1…
I hope these materials can be helpful to others who are considering pursuing an academic career!
4
1,304
Zu retweeted
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro
Whoever drives this at GEAR deserves a big round of applause, bullish on N2
We built a robotics data engine with Astra-level performance at 125x lower cost. And we are open-sourcing the weights today! Introducing Flora, a suite of open models to parse the physical world and turn raw footage into rich training signals.
Flora already powers our in-house data engine, labeling 200K hours of videos a day on 500 GPUs. We created it to scale the frontier for robot learning. Now you own the weights too.
Flora gives you:
- High-quality, timestamped captions to curate and filter human egocentric and robot data;
- Dense progress rewards for quality control and RL post-training;
- Rubrics scoring for real-world benchmarking
Try it today, links in thread:
Jensen keeps winning
NVIDIA has just announced a huge investment package into French startups.
@nvidia has now committed nearly $10 billion to the European tech ecosystem.
The French startups it is investing in include:
> Doctolib
> Entalpic
> Genesis AI
> Pigment
The company had already invested in other French companies such as Alice & Bob, Gradium and SCINTIL Photonics.
This is amazing news for the French and European ecosystem.
LETS GO
5
652
i love writing. i never let agents write for me. whether it's docs, or tweets, or shitposts, or even prompts (okay maybe that one's harder as agents manage other agents)
there's a simple reason for that: i still strongly believe that writing is thinking. it's a skill i hope to never give up. it's also important to me that whatever i say is always authentic, mistakes and all. my favorite posts are the ones that i just full send without any editing
i suspect it also helps with prompting. being able to clearly express what i want, and knowing the vocabulary i need to be able to direct my agents does seem to make for more effective usage of agents. in the same way that engineers have an advantage when we prompt agents to write code, because we are intimately familiar with the words that make up the discipline
knowing the right word incantations to get an agent to produce a desired goal almost feels like spellcasting. as models get more capable, and that input box looks more and more like a magic box that can output anything, words become the interface between you and the genie
there may come a time when we give up the keyboard in exchange for brain-computer interfaces. but until then, i'm writing every word.
Zu retweeted
Perhaps counterintuitive, but this is fundamentally true.
IK is built on a simple model of the arm and doesn’t account for things like friction, arm dynamics, environment. With RL you can get reactivity, distill to choose the right IK branches to avoid singularities.
Bitter lesson much?
Replying to @xiaolonw @chris_j_paxton
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.






















