Research Scientist @NVIDIAAI ; Training humanoids ๐ค ; Reinforcement learning PhD @TechnionLive ๐ฎ๐ฑ. Views are my own.
Israel
Joined March 2012
- Tweets2.2K
- Following683
- Followers2.7K
- Likes6.3K
Pinned Tweet
Animation ๐ค Robotics
ProtoMotions GTC 2026 release โ bridging the gap between digital humans and real humanoid robots.
Train in simulation. Deploy on hardware. One framework, one codebase.
nvlabs.github.io/ProtoMotionโฆ
If we have a sparse planner, all we need is a more reliable goal-conditioned controller?
๐คRobots need world models for visual planning, but they don't need to imagine every frame.
Introducing DreamCan: Sparse Visual Planning for Long-Horizon Robot Manipulation.
Project: dreamcan.github.io/
Code: github.com/HeegerGao/DreamCaโฆ
Paper: arxiv.org/abs/2610.24981
Chen Tessler retweeted
This is also why it's silly that projects still get released with "Code (coming soon)". Just release the code!
It doesn't matter if it's ugly, no human is going to read it, it's for the agents.
Not releasing the code risks people re-implementing it badly, and not realizing how great your work is!
Chen Tessler retweeted
Our team at NVIDIA is looking for interns for 2027! We are working on a broad range of topics related to humanoids, motion generation, and RL.
Apply here: nvidia.wd5.myworkdayjobs.comโฆ
Also feel free to contact me with a bit of info about yourself.
The rate in which everyone is adopting and implementing this reinforces to me how broken the conference system is.
Wait a few months for 3 random people to hopefully give a decent read, or... Let the community judge?
Community validation is 100x more important and impactful
We found a way to scale humanoid control that reduces GPU-hours by more than 1000ร.
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/
๐คฏ
We found a way to scale humanoid control that reduces GPU-hours by more than 1000ร.
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/
Joints, camera, action! conditions the video model on a rig and co-generates both the video and the motion directly on the target robot, bypassing challenges such as pose estimation and retargeting.
๐ฌ Excited to share our paper: Joints, Camera, Action! (JCA)
JCA generates video & 3D motion together for any rig!
Video models generate motion you can watch. But what if you could use that motion to animate a 3D asset, with motion that responds to its surroundings?
research.nvidia.com/labs/parโฆ
๐งต
Looks less AI generated than the Tesla roadster doing a kickflip over the Tesla truck.
Can't believe that's real.
This just makes so much sense.
Very intuitive idea ๐คฉ
Excited to share ๐๐ฅ๐๐๐: a whole-body tracker for humanoids that can exert forces from diverse poses and adapt to varied terrains.
Humanoid whole-body trackers are typically trained to follow references recorded on flat ground, carrying nothing. As a result, they know ๐ป๐ฒ๐ถ๐๐ต๐ฒ๐ฟ ๐๐ต๐ฒ ๐ณ๐ผ๐ฟ๐ฐ๐ฒ๐ ๐๐ต๐ฒ ๐ฟ๐ผ๐ฏ๐ผ๐ ๐บ๐๐๐ ๐ฒ๐
๐ฐ๐ต๐ฎ๐ป๐ด๐ฒ ๐๐ถ๐๐ต ๐ผ๐ฏ๐ท๐ฒ๐ฐ๐๐ ๐ป๐ผ๐ฟ ๐ฐ๐ฎ๐ป ๐๐ต๐ฒ๐ ๐ฎ๐ฑ๐ฎ๐ฝ๐ ๐๐ผ ๐ฑ๐ถ๐๐ฒ๐ฟ๐๐ฒ ๐๐ฒ๐ฟ๐ฟ๐ฎ๐ถ๐ป๐. BRACE closes both gaps in a single tracker.
Our key innovation is a ๐ฟ๐ฒ๐ณ๐ฒ๐ฟ๐ฒ๐ป๐ฐ๐ฒ-๐ฎ๐ฑ๐ฎ๐ฝ๐๐ฎ๐๐ถ๐ผ๐ป ๐ณ๐ผ๐ฟ๐บ๐๐น๐ฎ๐๐ถ๐ผ๐ป. Given humanoid reference trajectories, we adapt them to the terrain and the commanded force, then train the controller on these adapted references. Tracking them naturally produces the stance and bracing needed to handle the terrain and exert the force (or compensate for external loads). ๐ง๐ต๐ฒ ๐ฟ๐ฒ๐๐๐น๐๐ถ๐ป๐ด ๐ฐ๐ผ๐ป๐๐ฟ๐ผ๐น๐น๐ฒ๐ฟ ๐ฑ๐ฒ๐ฝ๐น๐ผ๐๐ ๐ฑ๐ถ๐ฟ๐ฒ๐ฐ๐๐น๐ ๐ผ๐ป ๐๐ต๐ฒ ๐ฟ๐ฒ๐ฎ๐น ๐ฟ๐ผ๐ฏ๐ผ๐.
To know more, please visit our:
Project page: multiplylabs.github.io/braceโฆ
Interactive demo: multiplylabs.github.io/robogโฆ
More details in this thread๐งต
1/n
Chen Tessler retweeted
We gave Claude and GPT a phone video of a robot wiping. In an hour, each built a full-dynamics digital twin.
Introducing RoboSimStudio ๐ค: an open and interactive toolkit for Agentic Robotics, enabling Fable, Astra and Argon AI agents to model sponges ๐งฝ, cables ๐งต and cloth ๐.
I disagree.
Maintaining decent code means your agent can keep building a larger Lego tower, composing together ideas.
Doesn't mean you have to code review it. There's many different ways. But crappy code is crappy code even in the agentic era.
my hypothesis is that maintainability is not as important as it used to be. the whole point of โmaking code easy to changeโ was because it took a lot of human effort to change code. well, agents make it trivial!
thought experiment: if the code is correct, bug free, fast and efficient with resources, and the agent is always easily able to change it, why does it matter?
not to mention that agents will continue getting more and more capable. my argument is that the definition of Good Code has changed
Chen Tessler retweeted
Day 276: Motion Dataset Prep (part 3) ๐
First 50k motions exported. 46k accepted, 2k rejected, and 2k in review. Excellent attrition.
This run sampled equally from the entire weighted motion/terrain matrix manifest. So the next longer motion-gen run won't have to redo prior work.
I had astra build me a manual review system where I can scrub through potentially bad motions and leave feedback at specific frames.
I've been using this for about a week now and it has been an absolute blessing.
We live in a world where you are only a prompt away from a workflow defining tool. What a magical time to be alive.
Feedback delivered, astra is on the fixes. โจ
Chen Tessler retweeted
Day 274: Motion Dataset Prep ๐
Rough checks looking good, but first I need to take a test pass and render out permutations of:
- speed
- approach angles / distances / facing direction
- terrain type (height, width, slope)
- general locomotion
- and a long list more...
I've tested separately as I worked on this, but it's likely there are some hidden issues.
If all that goes well, then i'll get started exporting the full dataset. Aiming for around 200gb in size. Max diversity.
Chen Tessler retweeted
Meet DyRAD: Radar NVS for dynamic driving scenes.
๐ก๐ Radar directly senses motion through Doppler, but prior radar NVS methods ignore this signal where it matters most: dynamic scenes.
๐ dyrad-nvs.github.io
๐ arxiv.org/abs/2609.39841
๐ป github.com/Dyrad-NVS/DyRAD
๐งต
Chen Tessler retweeted
#HomeBody is soon going to be public for everyone!
Make sure you star github.com/Stanford-TML/homeโฆ and follow the updates.
@giohuh_ @CaydenGu @TakaraTruong
Weโve updated the HomeBody repo with our code release schedule for the coming weeks ๐
!
Stay tuned here: github.com/Stanford-TML/homeโฆ
My experience with Astra is that it just needs a cheerleader.
I just tell it it did a good job and I believe it can do even better.
Then it goes and... does even better.
Chen Tessler retweeted
๐ค Ever wondered if you can train end-to-end humanoid policies entirely within a differentiable simulator?
๐ฅ Introducing ๐๐๐ป๐ฑ๐น๐ฒ๐ฑ ๐๐ผ๐ป๐๐ฎ๐ฐ๐ ๐๐ฟ๐ฎ๐ฑ๐ถ๐ฒ๐ป๐๐ (๐๐๐)
๐: bundledcontactgradients.githโฆ
๐: arxiv.org/abs/2609.30951
๐งต๐
Presented amazingly well. Highly recommend listening to the whole talk.
It's pencils down, people. Writing code by hand is no longer an economically viable skill for most programmers at most companies. But the future of making software has never been brighter. Don't you dare black pill this beautiful moment! youtu.be/vDjW_dRyKXY?si=6Fsfโฆ
Chen Tessler retweeted
A new version of SOMA Retargeter is live, making it even easier to bring motions generated by Kimodo and ARDY to your robot!
The new Robot Configurator feature helps create a reusable setup for retargeting motions to new humanoids you want to use.
github.com/NVIDIA/soma-retarโฆ
