Chen Tessler
@ChenTessler
Followers
1K
Following
4K
Statuses
958
Research Scientist at @NVIDIAAI, training simulated robots 🤖 with reinforcement learning. PhD from @TechnionLive 🇮🇱. Views are my own.
Israel
Joined March 2012
Excited to share our latest work! 🤩 Masked Mimic 🥷: Unified Physics-Based Character Control Through Masked Motion Inpainting Project page: with: Yunrong (Kelly) Guo, @ofirnabati, @GalChechik and @xbpeng4. @SIGGRAPHAsia (ACM TOG). 1/ Read along! 😃
9
91
377
My point was that there are no world models. Imo it's just a marketing stunt to make papers sounds shiny. It's like foundation models. They are simply generative models. But we needed a cool name to differentiate them from the "classic" generative models. One day we'll build something even better and we'll need to invent an even more nuanced term than world/foundation models 😆
1
0
1
RT @AbermanKfir: We discovered that imposing a spatio-temporal weight space via LoRAs on DIT-based video models unlocks powerful customizat…
0
38
0
RT @RL_Conference: 🚨🚨 RLC deadline has been extended by a week! Abstract deadline is Feb. 21 with a paper deadline of Feb. 28 🚨🚨. Please s…
0
19
0
Yep, basically my point here. I believe advances in sim make it much easier to build more complex scenarios. It used to be really hard with old simulators to build complex things. For example, IsaacGym with multiple objects -- just the tensor allocation logic was a nightmare, you never even reach the challenging part of coding the task itself 🥲 I think it's becoming easier and will continue to improve as people realize what use cases are interesting and build the tooling around it to accelerate research.
I'm not sure it's just about domain randomization. There used to be stability issues and physics artifacts. Think that's better now. It's hard to build complex scenes and code rewards etc... Locomotion is very easy in that regard. Single robot, only scene is terrain, rewards obs etc are very simple to get. Makes sense we'll first see a lot of progress in "locomotion" before we see similar advances in manipulation.
0
0
1
I'm not sure it's just about domain randomization. There used to be stability issues and physics artifacts. Think that's better now. It's hard to build complex scenes and code rewards etc... Locomotion is very easy in that regard. Single robot, only scene is terrain, rewards obs etc are very simple to get. Makes sense we'll first see a lot of progress in "locomotion" before we see similar advances in manipulation.
1
0
1
@AvivTamar1 First link is manipulation
Next step in dynamic dexterous grasping from NVIDIA: DextrAH-RGB! No more depth. We’re now consuming RGB stereo pairs, and the resulting perceptual system is much more robust. Trained entirely in sim (IsaacLab), leveraging fast tiled rendering, and deployed zero-shot to real.
1
0
0
@AvivTamar1 Not products, but impressive (imo) sim to real.
🚀 Can we make a humanoid move like Cristiano Ronaldo, LeBron James and Kobe Byrant? YES! 🤖 Introducing ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills Website: Code:
1
0
1
@rohan565singh This is what I'm using. All my code is open source. Releasing V2 soon with some big improvements
1
1
2
@rohan565singh It's good at discriminating but it doesn't hit binary labels. So you get (let's say..) 0.4 for fake and 0.6 for real. That's thanks to the gradient penalty regularizer. In terms of rewards I see ~0.2. I think as the policy improves the discriminator does too.
1
0
0
@rohan565singh In terms of accuracy, if we split on 0.5, it's able to separate real from fake almost perfectly (99.8% of the time).
1
0
0
Direct control from video 🤯
Next step in dynamic dexterous grasping from NVIDIA: DextrAH-RGB! No more depth. We’re now consuming RGB stereo pairs, and the resulting perceptual system is much more robust. Trained entirely in sim (IsaacLab), leveraging fast tiled rendering, and deployed zero-shot to real.
0
0
3
I feel this feature could have saved me from many bugs along the way. Can I get early access?
3
0
6
This is amazing! Great work @EugeneVinitsky !
We've built a simulated driving agent that we trained on 1.6 billion km of driving with no human data. It is SOTA on every planning benchmark we tried. In self-play, it goes 20 years between collisions.
1
0
8
@Rawlala1 @rohan565singh Yeah worth testing. Might check it later, if my code refactor tests pass without issues 🙈
1
0
1
@Rawlala1 @rohan565singh If it's successfully reconstructing the full motion then the average state distribution will ~ match randomly sampling poses from the data. No? Think just need to make sure not to terminate early unless it failed.
1
0
1