Dhruv Batra is Yutori’s cofounder and Chief Scientist and an adjunct professor at Georgia Tech. His research connects computer vision, embodied AI and models that operate web interfaces.[1][2]
Start with Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization →
Education
Links
Website Scholar
X · DhruvBatra_ Cofounder and Chief Scientist · Yutori
Adjunct Professor · Georgia Institute of Technology
Career
Batra previously led FAIR’s Embodied AI team as a Senior Director at Meta and held a tenured associate professorship at Georgia Tech. Earlier academic appointments included Virginia Tech and the Toyota Technological Institute at Chicago. His degrees are from Carnegie Mellon and the Institute of Technology at Banaras Hindu University.[1][3][4]
Contributions
Localize the image evidence behind a prediction
Grad-CAM uses gradients for a target concept flowing into the final convolutional layer to weight feature maps and produce a coarse localization map. This can show which image regions contributed to a classification, captioning or visual-question-answering result. Batra coauthored the method with researchers including Ramprasaath Selvaraju. A heatmap helps inspect model behavior; it does not by itself establish that a prediction is correct.[5]
Combine rendered interfaces with structured page access
At Yutori, Batra works on the Navigator model family. The n1.5 release combines visual interaction with DOM inspection and JavaScript execution. The model can fill multiple form fields or gather a dense page’s information in one structured step, then verify the result against the rendered interface. The release separates human and automatic evaluation tracks when reporting browser-task performance.[2][6]
Tools, writing & talks
Writing, research & talks
Start here
The original paper on gradient-weighted visual explanations for convolutional networks.
Start with the paper’s localization diagrams and formulation. Basic neural-network and gradient knowledge helps; compare the heatmap with the original input and prediction before treating it as an explanation.
Coauthor[5]Yutori's Dhruv Batra argues the web will not meet agents with APIs, because the long tail of 200 million sites will not ship MCP servers, and reading HTML does not save you when a score arrives later as JSON and sold out is expressed as a quantity of zero.
Catalog workOriginal source
Tools
- Navigator n1.5yutori.com
Yutori’s hybrid visual, DOM and JavaScript computer-use model release.
Cofounder and Chief Scientist of the publishing team[6]