Atharva Mete

AI Intern @agilityrobotics MS student at @ICatGT
Previous @iitbombay
Teaching robots how to learn.

[Atharva Mete @MeteAtharva]

Excited to share our #NeurIPS2024 work on QueST: Self-Supervised Skill Abstractions for Learning Continuous Control🦾🤖

QueST is a multitask latent-variable model that learns sharable low-level skills and outperforms Diffusion Policy, ACT and VQ-BeT by >13% in 5-shot transfer.🧵

[Atharva Mete @MeteAtharva]

@splicewiring @atlasdiscovery0 @ShaamilKarim1 @Cgensbigler Congrats on the launch!🥳

[Atlas Discovery @atlasdiscovery0]

Nine out of ten drugs fail in clinical trials. The reason isn't a shortage of data. It's that none of it connects.

We started Atlas Discovery to fix that with foundation models of patient drug response. We're backed by @ycombinator, @pearvc, and more.

[Atharva Mete @MeteAtharva]

@ryan_punamiya @AnthropicAI So far my every prompt to fable has been redirected to opus :(

[Shubham Sharma @HappyyPablo]

open sourcing Marlin-2B 🐟
a tiny VLM to extract structured information from videos

Marlin is finetuned for two questions devs want to ask in their videos: what is happening, and when?

Best open model in its weight class, competitive with Gemini-2.5-flash at only 2B params 🧵

[Atharva Mete @MeteAtharva]

Using human data in post-training was not in my 2025 bingo card
https://t.co/J7qhZ7IR9c

[Physical Intelligence @physical_int]

If we use our full pre-trained pi05 model, simply finetuning with human video data can double the performance on tasks that are depicted in the human videos!

[Atharva Mete @MeteAtharva]

@hausman_k Super cool! You think it can learn to do a fern latte art?

[Atharva Mete @MeteAtharva]

Noticed Neo at Times Square @1x_tech

[Ryan Punamiya @ryan_punamiya]

Robots struggle to learn new skills from human videos.

Why? We found that naive co-training produces disjoint distributions.

Our EgoBridge (NeurIPS’25) extends Optimal Transport to align human-robot latents, improving success by 44% and generalization to human-only tasks!🧵

[Jonah Siekmann @JonahSiekmann]

This is what I have been building for the last few months: a whole-body controller trained in Isaac Sim using RL. The ability to put the hands wherever they are needed and let the NN take steps and adjust stance as needed is a huge boost in capability for teleoperators + LfD.

[Albert Wilcox @albertwilcoxiii]

Imitation learning has seen great success, but IL policies still struggle with OOD observations

We designed a 3D backbone, Adapt3R, that can combine with your favorite IL algorithm to enable zero-shot generalization to unseen embodiments and camera viewpoints!

[Jeremy Collins @jerthesquare_]

Robotics data is expensive and slow to collect.

Robotics labs and companies spend months just to collect around 10k hours of demonstration data, all while that much video is uploaded to YouTube every 20 minutes. However, none of this video data contains action labels. How can we bridge the gap?

AMPLIFY solves this problem by learning Actionless Motion Priors that unlock better sample efficiency, generalization, and scaling for robot learning.

[Jack D. Carson @mtlushan]

neurips deadline is in 14 days. do I have a paper? no. do I have an abstract? no. do i have compute? not even a little bit. do I have data? sorta. do I have "the self assurance that [I will] succeed where lesser men failed"? you bet your ass I do.

[Shubham Sharma @HappyyPablo]

Tired of scrubbing through hours of footage to find the perfect clip? Shikai’s got your back! We’re building the world’s most *flexible* video search engine, all in open source @lossfunk !! 🚀✨

Check out how we did it on the recent @Project_11A video👇🧵

[Atharva Mete @MeteAtharva]

@sainingxie What do you think about compute not adaptive part? Vae-based compressors usually have a fixed compression ratio

[Atharva Mete @MeteAtharva]

@xiao_ted Very cool results! Though I wish there would have been some model architecture details in the report.

[Atharva Mete @MeteAtharva]

@tszzl https://t.co/sXfapIKhvi
Francois doesn't think so

[François Chollet @fchollet]

@NielsRogge This is the point of the training set, though -- to train your model on it. It would be even more impressive if the model had no prior exposure to ARC data, but the fact that the model was ARC-adapted via the training set absolutely does not invalidate its score.

[Atharva Mete @MeteAtharva]

@xiao_ted Is there any work that studies the "right kinds of data" problem?

[Simar Kareer @simar_kareer]

Introducing EgoMimic - just wear a pair of Project Aria @meta_aria smart glasses 👓 to scale up your imitation learning datasets!

Check out what our robot can do.

A thread below👇

[Atharva Mete @MeteAtharva]

@ARX_Zhang Is it compliant / back-drivable?

[Atharva Mete @MeteAtharva]

@MarcRigter Pretty cool work!
Compared to Genie-type models, I wonder if we can expect this to allow for better controlled/constrained generation?