Roei Herzig
@roeiherzig
Followers
1K
Following
8K
Media
120
Statuses
2K
Researcher @IBMResearch. Postdoc @berkeley_ai. PhD @TelAvivUni. Working on Compositionality, Multimodal Foundation Models, and Structured Physical Intelligence.
Berkeley, CA
Joined March 2017
What happens when vision๐ค robotics meet? Happy to share our new work on Pretraining Robotic Foundational Models!๐ฅ ARM4R is an Autoregressive Robotic Model that leverages low-level 4D Representations learned from human video data to yield a better robotic model. @berkeley_ai๐
9
73
437
Big personal news: I've traded my "extraordinary alien"๐ฝ status for a much cooler one--just a very standard permanent resident! Yay!๐๐ฅณ My green card has officially arrived! ๐ ๐บ๐ธ๐ฝ
5
0
24
We are working on that direction. Stay tuned ๐
GPT-5 Example: โHow many fingers per hand?โ We need vision-centric VLM โ and even more, vision-action-centric VLA. Language is a powerful tool, but itโs hard to bend it to the needs of vision and robotics.
0
0
2
๐ Excited to speak at the Agentic AI Summit NOW! ๐Frontier Stage - Session 4: Foundations of Agents Iโll give a 5-min spotlight on my research: ๐บ๐๐๐๐๐๐๐๐๐
๐ท๐๐๐๐๐๐๐ ๐ฐ๐๐๐๐๐๐๐๐๐๐๐ โ and why it matters for the future of vision & robotics
Really excited for the Agentic AI Summit 2025 at @UCBerkeleyโ2K+ in-person attendees and ~10K online! Building on our 25K+ LLM Agents MOOC community, this is the premier global forum for advancing #AgenticAI. ๐ Livestream starts at 9:15 AM PT on August 2โtune in!
0
3
14
๐ Excited to speak tomorrow at the Agentic AI Summit! Iโll give a 5-min spotlight on my research: ๐บ๐๐๐๐๐๐๐๐๐
๐ท๐๐๐๐๐๐๐ ๐ฐ๐๐๐๐๐๐๐๐๐๐๐ โ and why it matters for the future of vision & robotics ๐ค๐ง ๐ฏ https://t.co/4Fyv90sThN
#AgenticAI #AI #Robotics
0
1
5
ืื ืืืืื ืื ื ืืื. ืืชืงืจืช ืืืืืืช ืงืืืืช ืืื ืืืงืจื ืงืฆื, ืฉืืื ืืื ืืื ืืืื ืื ื ืืกืคืืง ืืืืื. ืืื ืื ืืืืืจ ืืืขืืืช ืฉืื ื ืืชื ืืืฉืื ืืื ืืืื (ืืืฉื ืจืืืืืื), ืื ืืื ืื ืืืืืจ ืขื ืืขืืืช ืืจืืฉ ืฉืืขืืื ืื ืืืื ืื ื ืืื ืืืืืืช. ืืืฉื ืืื ืจื ืืืคืืฉืืช ื ืืืจื ืืืืืืื ืก.
ืืขืืจ ืืื ืืคืื ืงืฆืื ืืืืช ืืืืืืื, ืขื ืืืื ืืงืกื, ืฉื ืืืืื ืืืืืื/ืชืืื ื ืืืื ืฉืื ืื ืืื ืืืืื ืืจืืฉ ืื ืื, ืืื ืืืขืืจ ืืจืฉืชืืช ืงืื ืืืืืฆืื (CNNืื) ืืืืืื ืชืืื ื ืืืืกืกื ืืจื ืกืคืืจืืจืื, ืฉืืชืืกืกื ืขื ืืืชื ืืืืจ ืืืกืืืจื ื-2017. ืืกืืฃ, ืื ืฉืฆืจืื ืืืืื ืื ืฉืืืจื ืืงืืืืช ืืืื AI ืืืืชื ืืกืืื. ืืฉืื ืืกืืื, ืืื
1
0
1
Awesome idea! Congrats folks๐
Excited to share Flow Matching Policy Gradients: expressive RL policies trained from rewards using flow matching. Itโs an easy, drop-in replacement for Gaussian PPO on control tasks.
0
0
4
My son?๐
0
0
1
Thanks @IlirAliu_ for highlighting our work!๐ ๐ Project page: https://t.co/gmPJqMLGNE ๐ Code: https://t.co/7XgQVSPDhI More exciting projects on the wayโstay tuned!๐ค
github.com
Official Implementation of ARM4R ICML 2025. Contribute to Dantong88/arm4r development by creating an account on GitHub.
Robots usually need tons of labeled data to learn precise actions. What if they could learn control skills directly from human videosโฆ no labels needed? Robotics pretraining just took a BIG jump forward. A new Autoregressive Robotic Model, learns low-level 4D representations
0
1
7
๐ Our code for ARM4R is now released! Check it out here ๐
github.com
Official Implementation of ARM4R ICML 2025. Contribute to Dantong88/arm4r development by creating an account on GitHub.
What happens when vision๐ค robotics meet? Happy to share our new work on Pretraining Robotic Foundational Models!๐ฅ ARM4R is an Autoregressive Robotic Model that leverages low-level 4D Representations learned from human video data to yield a better robotic model. @berkeley_ai๐
0
14
70
๐ Excited that our ARM4R paper will be presented next week at #ICML2025! If youโre into 4D representations and robotic particle-based representations, donโt miss it! ๐คโจ I wonโt be there in person, but make sure to stop by and chat with Yuvan! ๐
What happens when vision๐ค robotics meet? Happy to share our new work on Pretraining Robotic Foundational Models!๐ฅ ARM4R is an Autoregressive Robotic Model that leverages low-level 4D Representations learned from human video data to yield a better robotic model. @berkeley_ai๐
1
13
60
Love the core message here! Predictions โ World Models. Predictions are task-specific, but world models can generalize across many tasks.
Can an AI model predict perfectly and still have a terrible world model? What would that even mean? Our new ICML paper formalizes these questions One result tells the story: A transformer trained on 10M solar systems nails planetary orbits. But it botches gravitational laws ๐งต
1
0
3
Final chances to #ICCV2025 ๐ด๐บ Submit your best work to the MMFM @ ICCV workshop on all things multimodal: vision, language, audio and more. ๐๏ธ Deadline: July 1 ๐
openreview.net
Welcome to the OpenReview homepage for ICCV 2025 Workshop MMFM4
๐จ Rough luck with your #ICCV2025 submission? Weโre organizing the 4th Workshop on Whatโs Next in Multimodal Foundation Models at @ICCVConference in Honolulu ๐บ๐ด Send us your work on vision, language, audio & more! ๐๏ธ Deadline: July 1, 2025 ๐
0
1
4
๐จ Rough luck with your #ICCV2025 submission? Weโre organizing the 4th Workshop on Whatโs Next in Multimodal Foundation Models at @ICCVConference in Honolulu ๐บ๐ด Send us your work on vision, language, audio & more! ๐๏ธ Deadline: July 1, 2025 ๐
sites.google.com
News: ๐ Submissions are on OpenReview (Open Now!): https://openreview.net/group?id=thecvf.com/ICCV/2025/Workshop/MMFM4 MMFM has two paper tracks. More information under Call for Papers. Proceedings...
2
8
24
Re "vision researchers move to robotics"-theyโre just returning to the fieldโs roots. Computer vision began as "robotic vision", focused on agents perceiving & interacting within the world. The shift to "internet vision" came later with the rise of online 2D data. Org CV book๐
โwhy are so many vision / learning researchers moving to robotics?โ Keynote from @trevordarrell #RSS2025
0
0
6
Overall, I think the move from CMT to OpenReview was a great decision. Now if only we can improve the paper-reviewer matching system!
2
4
20
๐ Excited to share that our latest work on Sparse Attention Vectors (SAVs) has been accepted to @ICCVConference โ see you all in Hawaii! ๐ธ๐ด ๐ SAVs is a finetuning-free method leveraging sparse attention heads in LMMs as powerful representations for VL classification tasks.
๐ฏ Introducing Sparse Attention Vectors (SAVs): A breakthrough method for extracting powerful multimodal features from Large Multimodal Models (LMMs). SAVs enable SOTA performance on discriminative vision-language tasks (classification, safety alignment, etc.)! Links in replies!
0
5
32
Yes!๐ฅณ
It was nice engaging with the CV community on ways to stand out in the crowd. My answer was simple: work on robotics. There are so many unanswered problems and open pastures for research if you are a new researcher. Below are 6 problems I focussed on in my talk.
1
0
0
Honored to be named an ๐๐ฎ๐ญ๐ฌ๐ญ๐๐ง๐๐ข๐ง๐ ๐๐๐ฏ๐ข๐๐ฐ๐๐ซ for ๐๐๐๐ ๐๐๐๐ !๐ Grateful to contribute to the community and support the high standards of the conference. Maybe it's time to start thinking about AC-ing? ๐ #CVPR2025 @CVPR
1
0
12
Got a big, bold question? Let me know! Open to your questionsโambitious ones especially! ๐ค๐ฌ
๐จ Our panel kicks off at 11:30 AM in Room 207 AโD (Level 2)! Don't miss an amazing discussion with: Ludwig Schmidt, Andrew Owens, Arsha Nagrani, and Ani Kembhavi ๐ฅ
0
1
4