Shubham Tulsiani

@shubhtuls

Assistant Professor in the Robotics Institute, Carnegie Mellon University I want to build perception systems that can understand the physical world

shubhtuls.github.io

Joined January 2011

192 Following

3,587 Followers

Replies Reposts Media

Pinned
Shubham Tulsiani
@shubhtuls
Feb 27
[1/N] Current visual geometry prediction models primarily rely on labeled 3D data. Our CVPR26 paper, Flow3r, allows additionally leveraging unlabeled videos (using flow supervision) for scalable visual geometry learning, enabling accurate multi-view 3D reconstruction in-the-wild.
00:00
3
Shubham Tulsiani
@shubhtuls
Sep 11
We have seen impressive long-range 3D reconstruction systems, but 4D reconstruction remains limited to a few frames. Our Point4D framework enables this by one simple trick -- querying 3D points instead of 2D pixels to allow effective chaining for long range 4D reconstruction.
Minsik Jeon
@Minsik_Je0n
Sep 10
[1/N] Introducing Point4D: dense 4D reconstruction across hundreds of frames from a monocular video. Unlike existing feed-forward 4D methods that operate over short windows, Point4D keeps tracking in 3D, even across occlusions and out-of-frame motion.
00:00
1
Shubham Tulsiani
@shubhtuls
Aug 5
Large-scale flow matching policies have low reactivity (open-loop action execution within a chunk) and high perception-to-action latency (VLM inference + multi-step diffusion inference). We show how to make them useful for real-time reactive control. pi-r2-flow.github.io
Sungjae Park
@sungj1026
Aug 5
[1/N] Large flow policies (e.g., VLAs) have become extremely popular, unlocking scalability. However, they come at a cost: the robot cannot react quickly. π𝐑² enables them to achieve reactive, real-time control.
00:00
1
Shubham Tulsiani
@shubhtuls
Jul 9
RayRope is accepted to ECCV :) We added more experiments in the final version to show RayRope can be easily added to existing architectures across diverse tasks (e.g. EscherNet, TokenGS, UniMatch) and consistently improves performance.
Shubham Tulsiani
@shubhtuls
Feb 5
00:41
[1/N] Rotary Position Embeddings (RoPE) are ubiquitous across transformers that process tokens from 1D, 2D, or 3D grids e.g. language, images, or videos. Our RayRoPE formulation extends these to multi-view transformers. Paper and code: rayrope.github.io
1
Shubham Tulsiani
@shubhtuls
Jun 23
We present Lift4D -- allowing dynamic 3D reconstruction of objects from monocular videos with large motions, occlusions, and deformations. See @yehonation's thread for details, and be sure to check out the gallery of interactive in-the-wild results on the project webpage :)
Yehonathan Litman
@yehonation
Jun 23
To be honest, training on handmade 4D asset datasets is a dead-end. Almost all 4D asset data is synthetic and diverse real data barely exists, so models trained on it struggle to reconstruct objects that deform, get occluded, and move freely about the scene. Our new work,
00:00
3