FieldAI’s Ali Agha: Robots Need Physics, Not Just More Data
Automated Podcast 42:07
285 views · 14 likes Watch on YouTube ↗
Robots can hallucinate too.
In a chatbot, a hallucination may be an incorrect answer. In the physical world, it can mean a machine making the wrong move around people, heavy equipment, or an active jobsite.
In this live episode of Automated from Automate in Chicago, Brian Heater speaks with Ali Agha, founder and CEO of FieldAI, about robot hallucinations and what it will take to make physical AI safer in unpredictable real-world environments.
Ali argues that more robot data alone will not solve the problem. FieldAI combines data-driven learning with physics and uncertainty, giving a robot a way to recognize unfamiliar conditions, understand when its confidence is dropping, and slow down before a bad decision becomes a dangerous one.
His route to this problem runs through Qualcomm, NASA JPL, the Mars helicopter, and the DARPA Subterranean Challenge. Those projects forced robots to operate without maps, GPS, reliable communication, or any guarantee that the environment would resemble the training data.
The conversation also gets into why everything runs directly on the robot without Wi-Fi, 5G, or a cloud connection, how real deployments can break robotics’ data chicken-and-egg problem, and why construction sites create some of the richest training data in physical AI.
Ali also explains why FieldAI is building intelligence that can work across many kinds of machines instead of committing to one humanoid or hardware platform. Its software is already operating across 34 robot embodiments, including multi-ton vehicles, quadrupeds, and humanoids.
If you want to understand what it takes to move physical AI beyond controlled demos and into environments where edge cases are the everyday reality, this is the conversation.
KEY MOMENTS
00:00 Why edge cases define physical AI
01:24 Introducing Ali Agha and FieldAI
03:44 From MIT robotics to Qualcomm Snapdragon
04:56 How the Mars helicopter brought Ali to NASA JPL
06:45 How DARPA Challenges helped build modern robotics
12:07 When edge cases stop being edge cases
12:43 Why construction is harder than self-driving
15:51 Generalization versus getting robots deployed
16:41 One robot brain for 34 embodiments
19:05 Moving beyond the standard transformer playbook
19:33 How physics can help stop robot hallucinations
22:23 Why FieldAI runs entirely on the robot
23:38 Why world models still need real-world data
23:57 Breaking the robot-data chicken-and-egg problem
26:28 What the industry gets wrong about physics
26:52 Why physics-first AI is not the same as simulation
30:05 Giving AI more than raw camera and LiDAR data
31:38 What darkness does to robot perception
32:04 How robots handle dust, fog, smoke, and low light
34:47 Can industrial robot learning transfer to the home?
35:50 Why construction data is never boring
38:27 Why physical AI companies are building hardware
39:07 Why FieldAI is staying robot-agnostic
41:27 Why generalization requires generality
Connect with Ali Agha
https://www.linkedin.com/in/ali-agha-7aa5212a
Learn more about FieldAI
https://www.fieldai.com/
Learn more about FieldAI’s Field Foundation Models
https://www.fieldai.com/news/fieldai-announces-over-400m-in-funds-raised-to-advance-embodied-ai-at-scale
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, and Spotify. Also subscribe to 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/
In a chatbot, a hallucination may be an incorrect answer. In the physical world, it can mean a machine making the wrong move around people, heavy equipment, or an active jobsite.
In this live episode of Automated from Automate in Chicago, Brian Heater speaks with Ali Agha, founder and CEO of FieldAI, about robot hallucinations and what it will take to make physical AI safer in unpredictable real-world environments.
Ali argues that more robot data alone will not solve the problem. FieldAI combines data-driven learning with physics and uncertainty, giving a robot a way to recognize unfamiliar conditions, understand when its confidence is dropping, and slow down before a bad decision becomes a dangerous one.
His route to this problem runs through Qualcomm, NASA JPL, the Mars helicopter, and the DARPA Subterranean Challenge. Those projects forced robots to operate without maps, GPS, reliable communication, or any guarantee that the environment would resemble the training data.
The conversation also gets into why everything runs directly on the robot without Wi-Fi, 5G, or a cloud connection, how real deployments can break robotics’ data chicken-and-egg problem, and why construction sites create some of the richest training data in physical AI.
Ali also explains why FieldAI is building intelligence that can work across many kinds of machines instead of committing to one humanoid or hardware platform. Its software is already operating across 34 robot embodiments, including multi-ton vehicles, quadrupeds, and humanoids.
If you want to understand what it takes to move physical AI beyond controlled demos and into environments where edge cases are the everyday reality, this is the conversation.
KEY MOMENTS
00:00 Why edge cases define physical AI
01:24 Introducing Ali Agha and FieldAI
03:44 From MIT robotics to Qualcomm Snapdragon
04:56 How the Mars helicopter brought Ali to NASA JPL
06:45 How DARPA Challenges helped build modern robotics
12:07 When edge cases stop being edge cases
12:43 Why construction is harder than self-driving
15:51 Generalization versus getting robots deployed
16:41 One robot brain for 34 embodiments
19:05 Moving beyond the standard transformer playbook
19:33 How physics can help stop robot hallucinations
22:23 Why FieldAI runs entirely on the robot
23:38 Why world models still need real-world data
23:57 Breaking the robot-data chicken-and-egg problem
26:28 What the industry gets wrong about physics
26:52 Why physics-first AI is not the same as simulation
30:05 Giving AI more than raw camera and LiDAR data
31:38 What darkness does to robot perception
32:04 How robots handle dust, fog, smoke, and low light
34:47 Can industrial robot learning transfer to the home?
35:50 Why construction data is never boring
38:27 Why physical AI companies are building hardware
39:07 Why FieldAI is staying robot-agnostic
41:27 Why generalization requires generality
Connect with Ali Agha
https://www.linkedin.com/in/ali-agha-7aa5212a
Learn more about FieldAI
https://www.fieldai.com/
Learn more about FieldAI’s Field Foundation Models
https://www.fieldai.com/news/fieldai-announces-over-400m-in-funds-raised-to-advance-embodied-ai-at-scale
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, and Spotify. Also subscribe to 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/
Category (YouTube): Science & Technology
Playback is via YouTube's official embedded player. Data from YouTube; Exumo is not affiliated with YouTube.