Dhruv Batra on Why AI Agents Are Robots of the Web

AI can talk about the world. Acting in it is a different challenge.

Dhruv Batra says many robot demos released after ChatGPT are still little more than a language model inside a physical wrapper. The conversation may sound intelligent, but the real test is whether the robot can navigate, manipulate objects, understand physical space, and adapt when the world changes.

In this episode of Automated, Brian Heater speaks with the Yutori co-founder and chief scientist about what AI still cannot do, why robots do not always need maps, and how ideas from embodied intelligence are now powering a new generation of web agents.

Dhruv explains what it means for an AI system to have a “belief,” why a rational system should never assign exactly zero probability, and how artificial intelligence has narrowed the meaning of “thinking” to something researchers can measure.

They also discuss Dhruv’s work on visual question answering and Grad-CAM, his years leading embodied AI research at Meta FAIR, the limits of sim-to-real transfer, and the question roboticists ask before accepting a result: “Did you touch a robot?”

Finally, Dhruv explains why he sees Yutori’s browser agents as “robots of the web.” They perceive websites through pixels, take actions without hard-coded instructions, and learn from a digital world that changes constantly.
This is a grounded look at the line between chatbots, AI agents, and real-world intelligence.

KEY MOMENTS
00:00 Why LLMs still struggle in the physical world
00:55 Can AI benefit from philosophy and the humanities?
03:01 What it means for an AI system to have beliefs
04:15 Why intelligent systems need multiple plausible answers
06:02 Three ways to understand probability
08:44 Epistemic uncertainty and the capital of France
12:07 Why a rational AI cannot assign zero probability
13:48 The danger of AI “suitcase words”
15:40 Why speaking raises expectations for robots
16:42 When a robot is only ChatGPT in a body
18:38 How AI narrowed the meaning of “thinking”
21:53 How new paradigms rewire researchers
23:33 The AI ideas that may endure
26:30 Why rapid AI progress leaves little room for outside ideas
29:23 How researchers learn to predict the next important question
31:40 Did machine learning make hard-coded robotics wasted work?
34:50 Rodney Brooks and robotics’ sunk-cost problem
37:58 A robot that navigates with no map or LiDAR
39:24 Why maps are both overcomplete and undercomplete
41:49 Dhruv’s path from academia and Meta to Yutori
43:40 Grad-CAM and understanding neural network decisions
46:35 Why chatbots felt underwhelming
46:54 The question roboticists always ask: Did you touch a robot?
47:45 Where sim-to-real transfer works and where it still fails
51:32 Why Yutori’s AI agents are robots of the web
53:31 Hard-coded automation versus AI agents
54:16 Why dynamic websites break traditional automation
56:02 How live websites make AI agents better
57:09 One last question: What is the capital of France?

Connect with Dhruv Batra
https://www.linkedin.com/in/dhruv-batra-dbatra/

Learn more about Dhruv Batra
https://dhruvbatra.com/

Learn more about Yutori
https://yutori.com/

We’d love to hear from you. Have thoughts or guest suggestions?
Reach us at podcast@automate.org

You can find the transcript and more episodes of Automated at automated.fm

Unlock full access to Automated and explore everything automation.
Subscribe today and leave a review on YouTube, Apple Podcasts, Spotify, and the Automated Newsletter.
https://www.youtube.com/@automatedpodcast
https://podcasts.apple.com/us/podcast/automated-with-brian-heater/id1837762221
https://open.spotify.com/show/60olq6brlBEIJWggx2fMR6
https://www.automate.org/automation/automated-newsletter

You can also find us on:
LinkedIn https://www.linkedin.com/showcase/automated-podcast-by-a3/
Instagram https://www.instagram.com/automatedpod/