Co-founder & CEO @getcartwheel. Previously, Google. Also Universal Patterns. Love animation, diagrams & explainers.
Joined March 2009
- Tweets1.1K
- Following694
- Followers2.6K
- Likes1.6K
Today we’re introducing Cartwheel Control!
Teaching robots to perform rich, expressive motion using the large scale kinematics data we’ve been building at Cartwheel.
We’re starting with three open policies for:
Unitree G1
@menloresearch Asimov-1
@RoboParty_DEV RP1
Each turns reference motion into coordinated whole-body movement in physics simulation—from subtle gestures to cartwheels and backflips.
I’m excited about where this leads: as our motion models improve, robots with capable controllers can learn an increasingly rich vocabulary of movement.
The policies are free to download on Hugging Face, with open simulation and visualization code.
Huge credit to the open ecosystem making humanoid robotics move this fast—BeyondMimic & SuperMimic, MuJoCo Warp, mjlab, Menlo Research, RoboParty, and Unitree.
Run any of the 3 policies in ~2 minutes:
Blog: getcartwheel.com/blog/cartwh…
Weights: huggingface.co/andrew-cartwh…
Code (Apache 2.0): github.com/Cartwhl/cartwheel…
Also, if you’re building a robot and want custom motion training data or custom control policies for your hardware — reach out. We’d love to build with you.
"Robot shrug your shoulders"
"Stand on one leg"
"Jog in place"
You can make a robot do pretty much any action or pose with @getcartwheel's Swing Choreo model.
#robotics #animation
Berkeley keeps publishing remarkable work forcectrl.github.io/
2
175
A world model with actual moving(ish) people!
Preeeetty cool @theworldlabs
Thanks @c_valenzuelab for nominating @getcartwheel as a generational company on the Emergent Startup list from @EricNewcomer!
Nice to be recognized for our kinematic world models that animate robots, games and movies 🤸
Sneak peek of our new face performance capture model. Pretty accurate head shape @getcartwheel !
#mocap #animation
A morning at our team offsite.
You move. Cartwheel retargets it to a robot.
We call it vibe robotics.
Making robots move has always been hard. We’d like that to stop being interesting.
1
4
364
Retargeting is finally getting its due in the academy!
yy-gx.github.io/GNR/
2
150
Soft objects are hard to make in simulation.
Impressive work from the @Meta team!
projectsuperdex.com/
143
We make decisions about video A LOT. Excited for these video structured decision models.
VideoJev
System One Model for Video Understanding and Interaction
paper: huggingface.co/papers/2610.1…
5
416
The leap in realism is what gets me.
Meet Swing Choreo, our new generative motion model at @getcartwheel .
Up to four people performing together, with nuanced interactions. Longer sequences of actions, much more faithful to your prompt. Expressive faces, hands with great finger motion, people of different sizes moving across ground at different heights.
For studios, this is the jump from individual actions for individual characters to whole scenes, with timing and interactions generated together.
For robotics teams, it opens up hours of varied, realistic human motion for simulation.
And with agents like Astra and Opus, you can generate entire, stylistically consistent motion libraries for multiple characters at a time.
A huge step forward. Very proud of this team.
#Astra #Opus #Claude #chatGPT #Robotics
The next generation of explainer graphics will be incredible
I asked Claude Opus 5.5 to explain how a rocket engine works by building an interactive Raptor 3 you can take apart in your browser.
Cut it open, follow the oxygen and the methane through both turbopumps, then throttle it and watch the shock diamonds move.
airsup.ai/rocket-engine
5
599
A tennis swing is a person AND a racket. Now we can capture both in 3D from ordinary video.
We’ve been working toward this for a long time at @getcartwheel. Still early, but I’m so proud of Andrew and this team.
What would you want to capture first?
Ah!! We've been trying to get this working for months.
Basically, we can now capture a person and the objects they're interacting with in 3D from regular video. Motion for both, plus approximate object shape.
I've been excited about this for a while because so much of what people do involves stuff. If you want a robot to learn how to use a tool, or to understand someone's tennis swing, you probably want to know what the tool or racket is doing too.
Same goes for props in animation and everyday tasks in occupational therapy. There's a lot here.
Still imperfect, but this week it started working well enough that I wanted to show you.
1
247
Introducing compute-optimal scaling laws for human motion generation!
Human motion generation has historically been niche and difficult to scale. Animators, roboticists, motion analysts, and game developers all see the potential, but it rarely lives up to the rigor required in production.
We have long believed this is because all prior systems have been fantastically data constrained. Until today, we haven't been able to prove that scaling would actually help.
Today we present our flagship research paper: Compute-Optimal Scaling Laws for Human Motion Generation. (link: github.com/Cartwhl/scaling-l…)
We are building the world's largest, high-quality human-motion dataset. We trained hundreds of models across varying scales. They match the Chinchilla scaling found in language.
This places motion as the fifth data modality we understand how to scale. Text, images, audio, and video all have scaling laws. None of them really reflect the physical world. 3D human motion, across time, does.
Both autoregressive and flow matching motion models scale predictably. We can now predict, before training a model, how good it will be.
Motion is no longer data constrained. It is now a predictable and scalable frontier for physical and digital intelligence. Our largest models are training now, and the gap between generated motion and the real physical world is closing fast.
Paper: github.com/Cartwhl/scaling-l…
Blog: getcartwheel.com/blog/scalin…
Gemini 3.7 Flash is excellent at agentic vision tasks. We've been working with it for some time now. It's fast. And cheap. @andrew_n_carr likes it! 🤸♀️
#Google #Gemini
Less than two weeks since our last update, and we've launched a new version of our perception model, Comic 4.3.
It's much much faster — now the fastest pro quality motion capture!
Best of all, it's just as fast for multiple characters as for a single character. And, you can batch process all your shots at the same time. So, 1 or 100 performances — super fast 🤸♂️
Also cool, you can use it directly in Unreal Engine, and soon Maya, Blender and Unity.
Live @getcartwheel.
We ran some shots through different tools and timed it all to see the speed.
#robotics #unreal #maya #blender #AI #unity #mocap #computervision
Feet on the ground.
It's always been one of the hardest problems in capturing human motion.
We've released an updated flagship perception model Comic 4.2.
We've focused on an advanced foot physics estimation system, properly estimating force, friction, toe roll, and elevation changes — all from simple video input.
For those making games and doing 3D animation, you can use it directly in Unreal Engine and soon Maya and Blender.
Mocap on tap, as we like to say. It's fast, accurate, and inexpensive 🤸♀️
Live at @getcartwheel. Here's a quick look!
#robotics #unreal #maya #blender #AI
@GoogleDeepMind's Gemini Robotics 2 is the most impressive robot movement demo I've seen. Really shows the promise of humanoid robot movement 🤸♀️
youtube.com/watch?v=9MNLEAzA…