Chen Tessler

Excited to share our latest work! 🤩 Masked Mimic 🥷: Unified Physics-Based Character Control Through Masked Motion Inpainting Project page: research.nvidia.com/labs/par/maske… with: Yunrong (Kelly) Guo, @ofirnabati, @GalChechik and @xbpeng4. @SIGGRAPHAsia (ACM TOG). 1/ Read…

Is there a way to mute all posts/replies tagging Grok? Twitter has become turbo stupid.

Want robot imitation learning to generalize to new tasks? Blindfold your human demonstrator! Best robotics paper at EXAIT Workshop #ICML2025 openreview.net/forum?id=zqfT2… Wait, why does this make sense? Read below!

Thrilled to share our #ICML2025 paper “The Courage to Stop: Overcoming Sunk Cost Fallacy in Deep RL”, led by Jiashun Liu and with other great collaborators! We teach RL agents when to quit wasting effort, boosting efficiency with our proposed method LEAST. Here's the story 🧵👇🏾

Heading to #ICML2025 to share our GBRL poster! • Gradient-boosted trees for RL • Strong performance & OOD robustness • CUDA-fast, SB3-ready, lightweight deployment Let’s talk in Vancouver! 📄 arxiv.org/pdf/2407.08250 💻 github.com/NVlabs/gbrl @DalalGal @ChenTessler #RL #AI

Check out Uri's work on combining RL with ComfyUI. Designing complex flows is a PITA and seems like it's another task that LLMs are pretty good at :)

Can we teach dexterous robot hands manipulation without human demos or hand-crafted rewards? Our key insight: Use Vision-Language Models (VLMs) to scaffold coarse motion plans, then train an RL agent to execute them with 3D keypoints as the interface. 1/7

🦋 I'm getting butterflies On Monday, we are launching a research preview of something amazing, bizarre, & ambitious at the same time. We can't wait for all of you to experience a glimpse of the future we have built with @CaggianoVitt & @MyolabAI team #proud

Behavioral Foundation Models (BFMs) trained with RL are secretly more powerful than we think. BFM’s directly output a policy believed to be near-optimal given any reward function. Our new work shows that they can actually do much better:

Public service announcement -- if you're making a new dataset of human motions in SMPL-X format using a marker-based system, please use MoSh. If you first compute a skeleton and then transfer this to SMPL-X, you will lose a lot of realism. MoSh fits SMPL-X to the markers…

Full body tracker now on a deployed G1 🤩

PPO is very frustrating. It simply works. Always. No matter what I throw at it. Other methods... not so much.

Thread packed with useful info and ideas! This stair walking behavior is really neat. I've always criticized RL controllers (mine included) that they lack proper spatial understanding. They are very robust, they can go up and down stairs without falling down, but it always feels…

Here’s our latest RL update: Natural Mogging (thread below!)

These results really surprised me. In some tasks where you can use decision trees, they exhibit superior performance and properties. One cool example is robustness to noisy or irrelevant features. NNs struggle ignoring these, while GBTs don't even flinch.

Will present some of our newest humanoid results!

Happy Throughput Thursday! We’re excited to release Tokasaurus: an LLM inference engine designed from the ground up for high-throughput workloads with large and small models. (Joint work with @achakravarthy01, @ryansehrlich, @EyubogluSabri, @brad19brown, @jshetaye,…

Also a reason why language commands aren't enough. This understanding is behind some design decisions we had in PDC -- zhengyiluo.com/PDC-Site/ .

1/6🚀 New #ACL2025Findings: We show you can predict if Chain-of-Thought (CoT) reasoning will succeed — before any tokens are generated! This works with LLMs not specifically trained for reasoning—meaning powerful signals emerge naturally in early processing.

Normalizing Flows (NFs) check all the boxes for RL: exact likelihoods (imitation learning), efficient sampling (real-time control), and variational inference (Q-learning)! Yet they are overlooked over more expensive and less flexible contemporaries like diffusion models. Are NFs…

excited to finally share on arxiv what we've known for a while now: All Embedding Models Learn The Same Thing embeddings from different models are SO similar that we can map between them based on structure alone. without *any* paired data feels like magic, but it's real:🧵