Dhruv Batra is the co-founder and chief scientist of Yutori, where he develops computer-use agents that navigate websites and perform practical tasks. His research has shaped visual question answering, interpretable neural networks, embodied intelligence, and browser automation.
Batra co-authored the original 2015 Visual Question Answering research, which challenged machines to answer natural-language questions about images. Later research examined whether these systems understood visual information or exploited patterns in the questions. The VQA challenge series received the Mark Everingham Prize in 2025.
He also co-developed Grad-CAM, which uses gradients to identify image regions influencing a neural network’s predictions, making decisions in classification, captioning, and visual question answering more interpretable. As an associate professor at Georgia Tech, he received the 2019 Presidential Early Career Award for Scientists and Engineers for research on explainable AI and neural-network interpretability.
His work on Habitat, an open platform for photorealistic 3D simulation, extended computer vision into environments where agents must navigate, follow instructions, and act. He subsequently led FAIR Embodied AI at Meta as a senior director, working on robotic navigation and manipulation, language-guided systems, and the multimodal assistant in Ray-Ban Meta smart glasses.
In March 2025, Batra, Devi Parikh, and Abhishek Das introduced Yutori and announced $15 million in seed funding. The company applies perception-and-action research to digital assistants that handle everyday online tasks.
What browser agents need to work
- The long tail of the web will not become API-ready. Restaurants publish menus as image galleries or scanned PDFs; school districts bury procurement details in institutional portals or Freedom of Information Act workflows. Batra argues that practical agents must operate on existing interfaces because these organizations are unlikely to rebuild their infrastructure around machine-readable endpoints.
- Rendered browser state is the source of truth. Initial HTML frequently omits information loaded asynchronously or produced through rendering logic. An item’s availability, for example, may appear only as a disabled or grayed-out option after inventory data has been processed. Visual grounding lets agents inspect the same interface people actually see.
- Navigator combines visual perception with programmable action. Yutori’s Navigator initially paired screenshots with browser actions such as clicking, typing, and scrolling. Navigator n1.5 adds generated JavaScript, allowing agents to complete multiple form fields simultaneously while checking their results against the rendered page.
- Agent evaluation must include accuracy, latency, and cost. Yutori reported 97.3% human-evaluated success for Navigator n1.5 on Online-Mind2Web and maintains Navi-Bench for tasks on live websites. Batra emphasizes that practical deployment also depends on per-step latency and total cost per completed task.
In a 2024 essay on the limitations of the term “large language model”, Batra described these systems as learners of symbol sequences while distinguishing their capabilities from the harder problems of visual perception and physical control. His browser agents pursue precisely that gap: turning systems that process information into systems that perceive an environment, act, and verify the outcome.
