Raktim Gautam Goswami

Huge congratulations to the entire team! Excited to see this vision come to life and looking forward to the breakthroughs in world models ahead. 💪🚀

@sainingxie @choang333 @ylecun @amilabs Congratulations!

i’m joining forces with @ylecun and an incredible group of people to start AMI Labs @amilabs. AMI isn’t a conventional lab. we don’t intend to become one. a lot to say about why this moment matters, but for now we’re heads down building. join us: amilabs.xyz

𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝗘𝗕-𝗝𝗘𝗣𝗔 ⚡ An open-source library making JEPAs accessible, trainable on a single GPU in hours! 🚀 🔗 Paper: arxiv.org/abs/2602.03604 💻 Code: github.com/facebookresear…

@DavidJFan @AIatMeta @ylecun Thank you, David! It was great working with you as well. I really appreciated all your support and guidance. I learned a lot working together. And yes, we’ll definitely have to keep the tradition going and try out some new restaurants next time 😄

@raktimgg @AIatMeta @ylecun Congrats Raktim!! It was a pleasure to have you here in the NYC office 😁 I learned so much from you about robotics and was really impressed with your attention to detail and persistence in getting things to work. Also enjoyed having some authentic Indian food :)

Just wrapped up my 7-month internship at Meta FAIR @AIatMeta. A deeply valuable experience, and I’m thankful for everything I learned along the way. During the internship, I worked on world models for dexterous manipulation under the guidance of @ylecun . (1/4)

Looking forward to continuing research in this domain and hopefully collaborating again in the future. Grateful for the journey and excited about what’s next. (4/4)

I’m also grateful to @DavidJFan, @GaoyueZhou, @JimmyTYYang1, @michaelrabbat, and the rest of the team for the collaboration and insightful feedback. Beyond the technical work, I grew greatly in how I communicate ideas, present my work, and think about research impact. (3/4)

As 2025 comes to an end, I feel thankful for these opportunities and humbled by how much there is still to learn. I’m excited to continue this journey in 2026 and hopefully contribute further to the AI and robotics community. (4/4)

None of this would have been possible without my co-authors, collaborators, and mentors; I’m deeply grateful for their support. I’m also thankful for the opportunity to intern at Meta-FAIR this past year; will share more about it in a separate post. (3/4)

Concluding 2025 on a grateful note. This past year, I had the privilege of presenting my work at top AI and robotics conferences: NeurIPS, CVPR, ICRA, IROS, and WACV. From robot parades to cutting-edge AI talks, each conference offered invaluable learning and inspiration. (1/4)

@tomchapin We haven’t released the code yet, which is why the link isn’t working. We’re working on getting it ready and hope to share it soon.

@raktimgg The code link in the project page doesn’t work?

Introducing DexWM: Dexterous Manipulation World Model tl;dr: train on human videos; fine-grained actions; hand-consistency loss; DexWM+MPC -> zero-shot dexterous manipulation w/ @_amirbar, @DavidJFan, @JimmyTYYang1, @GaoyueZhou, P. Krishnamurthy, M. Rabbat, F. Khorrami, @ylecun

@nazirjonov_ @_amirbar @DavidJFan @JimmyTYYang1 @GaoyueZhou @ylecun Task planning can be fairly sensitive to factors like the scale of encoder’s embedding space. While further tuning (e.g., emb. norm) may improve DINOv3 reuslts, we used the encoder off-the-shelf to show that DexWM is competitive with other encoders and not limited to DINOv2.

@raktimgg @_amirbar @DavidJFan @JimmyTYYang1 @GaoyueZhou @ylecun Great work, I found interesting that DINOv3 encoder performs worse compared to DINOv2 in task success rate. Any ideas why that might be the case?

@AlexiGlad It can indeed be expensive if we use too many frames. But in practice, we can cover a relatively long time span with small number of frames by using a low sampling freq. For eg, in our work we sample 8 frames from video sequences of ~4 seconds, and this already performs well.

@raktimgg got it, isnt that super expensive then? meaning your world model likely cannot span too many frames?

@AlexiGlad The sentence "observations ... features" refers to the diffusion policy baseline, where we only use average-pooled DINOv2 features. For DexWM, we use all patch tokens similar to DINO-WM. In my experience, pooled features work worse than full patch tokens for such world models.

"observations are encoded using the average-pooled DINOv2 patch features" just to clarify, are you guys predicting the average patch features, meaning you predict one patch per frame? did you ever try predicting the cls token instead of the average-pooled features? did you ever have issues because of only predicting a single average pooled feature rather than all the patch tokens like dinoWM?

Paper Title: World Models Can Leverage Human Videos for Dexterous Manipulation

Check out our paper and project page for more details: Project Page: raktimgg.github.io/dexwm/ Preprint: arxiv.org/pdf/2512.13644 Huge thanks to all my co-authors for their support and guidance throughout this project🙏