Rohun Tripathi

Rohun Tripathi is an AI researcher whose work spans computer vision, multimodal models and robotics. He holds computer science degrees from IIT Kanpur and Cornell Tech, and his early papers explored scarce-data segmentation, audio CAPTCHA usability and faster object detection. He coauthored SIDGAN, Molmo and PixMo, Molmo2 and MolmoBot. As of October 2026, he is a Research Scientist at Ai2, where he builds multimodal foundation models for vision, robotics and multimodal reasoning.

Now at: Allen Institute for AI (Ai2)
Works: 14
Citations: 34
h-index: 3

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

2018

Segmentation with scarce data

With a B.Tech. in computer science from IIT Kanpur and a master’s in computer science from Cornell Tech, he coauthors “Semantic Segmentation with Scarce Data” in 2018. The method combines coarse annotations with fine supervision to improve segmentation in scarce-label settings.
2019

Audio reCAPTCHA accessibility

In 2019, he coauthors an ACM ASSETS paper with Mohit Jain, Ishita Bhansali and Pratyush Kumar that introduces reCAPGen, a system using automatic speech recognition to generate secure audio CAPTCHAs. The study evaluates four schemes with 60 sighted and 19 visually impaired participants.
2020

ASAP-NMS object detection

In 2020, he coauthors ASAP-NMS, which speeds non-maximum suppression in object detectors. The paper reports NMS latency on a CPU falling from 13.6 milliseconds to 1.2 milliseconds, while preserving the accuracy of a state-of-the-art two-stage detector on COCO and VOC datasets.
2023

Video dubbing with SIDGAN

In 2023, he coauthors SIDGAN, which applies shift-invariant learning to generate high-resolution dubbed videos with synchronized lip motion. It preserves identity and scene details such as head pose and lighting.
2024

Molmo and PixMo

In 2024, he coauthors Molmo, an open family of vision-language models, and PixMo datasets for image captioning, question answering and 2D pointing. The paper makes model weights, datasets and source code available.
January 2026

Molmo2 video grounding

In January 2026, he coauthors Molmo2, an open family of vision-language models for video understanding and grounding. It introduces seven video datasets and two multi-image datasets, with point-driven grounding for single-image, multi-image and video tasks.
March 2026

MolmoBot robot manipulation

In March 2026, he coauthors MolmoBot, a robot manipulation system trained with 1.8 million expert trajectories generated in simulation. It reaches 79.2% success on real-world tabletop pick-and-place with zero-shot transfer and no real-world fine-tuning.
As of October 2026

Research scientist at Ai2

As of October 2026, he is a Research Scientist at Ai2, where he builds multimodal AI systems and foundation models for vision, robotics and multimodal reasoning.