Dhruv Shah

@shahdhruv_

professor @Princeton

San Francisco, CA

dhruvshah.me

Joined April 2012

1,329 Following

7,088 Followers

Replies Reposts Media

Dhruv Shah
@shahdhruv_
Oct 23
How should we train world models that are controllable and physically realistic? We show that maximizing "action influence" is an excellent objective for training high fidelity world models, dropping data requirements by 100x! princeton-prism.github.io/action-influen…
Samuel M Bateman
@smbml_
Oct 23
🚨 Stop feeding your world model the same boring demos. Switch to free-range play data! 🚨 Our CoRL 2026 paper finds we can optimize autonomous data collection to focus on missing action consequences, achieving up to 85x the data scaling efficiency of expert data collection!🧵
Dhruv Shah
@shahdhruv_
Oct 22
We find that a VLM can teach itself to understand language goals, reason about them, and command robot actions -- we call this Action Self-Distillation! ASD can perform language navigation by outperforming any VLA trained on hand-annotated data. asd-nav.github.io
Bowen Feng
@BowenFengbf
Oct 22
Can robots learn to reason and act with NO human-annotated language or reasoning labels? Introducing Action Self-Distillation (ASD), one model that teaches itself language goal-conditioned reasoning and navigation directly from robot demonstrations. #CoRL 2026🤖
3
Dhruv Shah
@shahdhruv_
Oct 15
We scaled up language-action pre-training to 10k hours across robots and egocentric videos. The key to facilitating motion transfer: using motion-level reasoning as a shared language across all the embodiments! ego-lap.github.io
Lihan Zha
@LihanZha
Oct 9
What’s missing from robot foundation models? An explicit understanding of WHY motions work: why move in this direction, why place the object in this way, how hard to grasp the object. This motion-level reasoning bridges instructions to actions—and provides knowledge that can
2
Dhruv Shah
@shahdhruv_
Sep 23
We scaled up whole-body navigation for humanoids (29-DoF joint-space actions) entirely using sim data and it can follow fine-grained instructions in complex real environments zero-shot! tango-vla.github.io Large-scale sim for training policies is really starting to work...
Anqi Li
@lnq105202
Sep 9
💡 Humanoid navigation is more than a 2D problem. How can a robot use its whole body to traverse cluttered 3D spaces? 💃🏻 Introducing our #CoRL2026 work TANGO, a whole-body VLA for humanoid navigation that directly predicts 29-DoF joint actions from RGB observations, enabling
2
Dhruv Shah
@shahdhruv_
Jul 30
One of my favorite parts of this release is our new on-device policy (GROD2), which incorporates cross-embodiment pre-training to seamlessly learn a generalist policy. This enables specializing to new embodiments in <200 demos without losing the pre-training goodness of Gemini.
Google DeepMind
@GoogleDeepMind
Jul 30
01:32
One brain. For any robot. 🤖 We’re launching Gemini Robotics 2: our next-generation physical AI bringing full body intelligence to humanoids, advanced dexterity, multi-robot teamwork and more.
6