22Astronauts

With Neil Nie, Co-founder & CEO of Verne Robotics.

Neil came to the US as a kid and grew up in a small town in rural Massachusetts, writing software mostly on his own. As a high school runner he built a workout tracking app for his coach when the first Apple Watch came out, and with a friend he built a self-driving golf cart.

He went on to do robot learning research with Shuran Song at Columbia and with Jiajun Wu in Stanford's SVL, spent about a year on Apple's fitness tracking team, and left his PhD at Berkeley to start Verne in summer 2025 with a friend from high school.

"If you have an arm that can only carry three or four kilograms, certain applications, especially in industrial settings, are simply not going to make sense."

Verne builds dual-arm mobile robots for manufacturing and logistics and owns the whole stack, from pre-training and post-training to the controllers on the robot. Customers today range from semiconductor supply chains to packing operations for contractors serving the US defense industry.

In this episode, Neil and Ilir talk about his mother's advice to turn down an Apple salary after college, why research still matters when AI can answer most questions, how Verne got into YC without a single customer, why he thinks industrial robots don't need legs, and why he wants customers to build on Verne's robots the way developers built on the App Store.

𝐒𝐏𝐎𝐍𝐒𝐎𝐑

This podcast is brought to you by CoreWeave⁩ providing the specialized cloud infrastructure that powers physical AI and large-scale model training.

Verne Robotics: https://vernerobotics.com

Neil Nie on X: https://x.com/neil_nie_

Neil Nie on LinkedIn: https://www.linkedin.com/in/neilnie

Ilir on X: https://x.com/IlirAliu_

Ilir on LinkedIn: https://www.linkedin.com/in/ilir-aliu

22Astronauts: https://22astronauts.vc/

Hosted by Ilir Aliu

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

0:00 What Verne Robotics builds

2:54 First-generation immigrant with entrepreneurial parents

3:43 Building a self-driving golf cart in high school

4:06 Columbia with Shuran Song, Stanford with Jiajun Wu

6:57 A workout app for his track coach

8:29 A year on Apple's fitness team

10:04 Alone in rural New England, then the robotics club

13:00 Should you still do a PhD when AI does research?

14:50 "Back then robotics just didn't work at all"

15:45 His mom's advice: skip the Cupertino salary

18:42 There are no shortcuts

23:03 Starting Verne and dropping out of the PhD

23:47 Why the company is named after Jules Verne

26:01 Robots that help businesses make and ship more

28:21 Selling into manufacturing as a first-time founder

30:58 Getting into YC with zero customers

33:20 The Pareto frontier of what robots should do

34:16 Windows XP vs. the App Store: customers as robot developers

38:14 The questions critics ask: vertical integration, deployment, form factor

41:54 Why legs are optional and payload is not

43:17 Design and public perception

45:00 What's coming next

In Ep #113 of 22Astronauts, Ilir Aliu sits down with Ali Ansari, Founder & CEO of ⁨Micro1.

Robotics teams still struggle to say which policy model is better, because almost nobody evaluates them on the same real hardware. Ali’s view is that measurement comes first and improvement follows. micro1 builds expert-generated training data and evaluations for most frontier AI labs, and by Ali’s account robotics is already above $100M of its run rate.

Ali was born in Iran, moved to the US at ten after his family won the green card lottery, and studied CS and math at Berkeley before a short stint at Stanford. micro1 began as an AI interviewer he built to screen engineers for his own software agency and is now valued at $4B.

“If you can precisely measure intelligence, you by definition have a reward model to train on.”

In this episode, Ali and Ilir cover why robot models will learn from human actions the way LLMs learned from human text, how micro1 collected 100,000+ hours of egocentric household video, and why cheap hardware with heavy annotation scales better than expensive sensors.

They also get into the new Malibu lab for safety evals on real arms and humanoids, how micro1 serves competing labs at the same time, and what Ali learned from the sacrifices his parents made.

𝐒𝐏𝐎𝐍𝐒𝐎𝐑

This podcast is brought to you by ⁨@CoreWeaveYT⁩ providing the specialized cloud infrastructure that powers physical AI and large-scale model training.

micro1: https://www.micro1.ai

Ali Ansari on X: https://x.com/aliansarinik

Ilir on X: https://x.com/IlirAliu_

Ilir on LinkedIn: https://www.linkedin.com/in/ilir-aliu

22Astronauts: https://22astronauts.vc/

Hosted by Ilir Aliu

TIMESTAMPS

0:00 What micro1 does and who Ali is

2:00 Born in Iran and winning the green card lottery

5:00 What the move cost his parents

9:40 Raising $100M+ at 25 and staying capital efficient

13:35 Why experts outside San Francisco change how people see AI

16:50 The two theses behind micro1’s move into robotics

21:40 From AI recruiter to training data for frontier labs

25:40 Synthetic vs. human data: simple hardware, heavy annotation

27:50 Why robotics needs physical benchmarks, not simulation

30:15 Inside the Malibu lab: safety evals and post-training

31:35 Serving competing labs without leaking information

34:00 Hiring, risk appetite and a role that changes every month

37:00 Mentors: Sean Rad and Joshua Browder

38:50 The Sunday reflection hour

In Ep #112 of 22Astronauts, Ilir Aliu sits down with Joe Harris, Founder & CEO of Alloy Robotics.

Before starting Alloy, Joe helped scale a consumer telehealth platform to a $1B+ acquisition. Now, he is bringing those high-growth scaling principles to physical AI.

"Reliability is measured at the action level. If every action is 99% accurate and you multiply that across 10,000 actions a day, your robot's actual uptime degrades incredibly fast."

Alloy Robotics builds the "AI Robotics Engineer", a specialized data and agent platform that ingests robot logs, telemetry, and sensor streams to automatically detect, diagnose, and resolve field failures.

In this episode, Joe and Ilir break down why general-purpose robotics hype collides with real-world fleet engineering, why hardware teams need SRE and agentic tooling, and how software remains the highest-leverage layer in robotics.

𝐒𝐏𝐎𝐍𝐒𝐎𝐑

This podcast is brought to you by ‪ ⁨@CoreWeaveYT⁩ providing the specialized cloud infrastructure that powers physical AI and large-scale model training.

Follow Joe Harris on X: https://x.com/_joe_harris_

Alloy Robotics: https://www.usealloy.ai/

22Astronauts: https://22astronauts.vc/

Hosted by Ilir Aliu

TIMESTAMPS

(00:00) Building the First AI Robotics Engineer

(04:05) Don't Sell a Platform, Sell a Solution

(09:38) Lessons from Scaling a $1B+ Business

(14:38) Scar Tissue vs. Intellectual Learning

(20:40) Why Hardware Needs SRE & Agentic Tooling

(27:58) How Platform Shifts Expand Total Market Cap

(34:25) The Hidden Cost of Robotics Log Data

(41:10) Global Supply Chains & Re-Industrialization

(51:12) The Compound Math of 99.999% Fleet Reliability

Rajat Bhageria is the Founder & CEO of Chef Robotics (San Francisco, CA), the leading physical AI company deploying AI-guided robot arms across food manufacturing and production facilities.

Instead of chasing restaurant cooking hype or $15/hr fast-casual makelines, Chef Robotics pivoted to high-volume food factories (including Amy's Kitchen, Chef Bombay, Gate Gourmet, and Sunbasket), recently surpassing 100M+ servings deployed in production across the US, Canada, and Europe.

After studying economics at Wharton and robotics at Penn's GRASP Lab, Rajat sold his first computer vision startup (ThirdEye) at age 20, co-founded pre-seed fund Prototype Capital, and launched Chef Robotics in 2019.

Backed by $65M+ in funding (from Kleiner Perkins, Avataar Ventures, Construct Capital, and SVB), Chef Robotics uses real-world production data to build domain-specific Food Foundation Models (FFM) and world models for deformable material handling.

𝐒𝐏𝐎𝐍𝐒𝐎𝐑This podcast is brought to you by  CoreWeave  providing the specialized cloud infrastructure that powers physical AI and large-scale model training.

Chef Robotics: https://chefrobotics.ai

Rajat on LinkedIn: https://www.linkedin.com/in/rajatbhageria/

Ilir on X: https://x.com/IlirAliu_

Ilir on LinkedIn: https://www.linkedin.com/in/ilir-aliu/

Timestamps:

0:00 Why Chef Robotics walked away from restaurant deals to focus on food factories

3:39 Growing up in Cincinnati: Reading biographies of Carnegie, Rockefeller & Edison

8:33 Building computer vision for the visually impaired (ThirdEye) at Penn

12:35 Basic research vs. Product execution: Why shipping matters

15:14 Why cooking robots fail: "Customers don't want cooking robots, they want assembly"

18:05 Expanding beyond prepared meals: Airline catering, meatpacking, and CPG kitting

22:04 Selling before building: Raising seed capital from Kleiner Perkins

25:34 Why every founder's primary job is persuasion and sales

33:23 The "Value Investing" approach in robotics: Finding high-volume factory assembly

36:08 Specialists vs. Humanoids: Why industrial plants need "Home Depot, not Costco"

41:34 Learning from the 2015 robotics graveyard: Why ROI trumps generalizability

46:28 Real-world production data vs. simulation: The Food Foundation Model

52:52 The Goal & Bottleneck Theory: Focusing strictly on operational constraints

59:20 Advice to my younger self: Category creation takes time

Almetra is turning the perception stack from self-driving cars into factory intelligence.

I sat down with Silviu Homoceanu, Co-Founder & CTO of Almetra, to talk about growing up with computers in post-communist Romania, eight years inside VW's self-driving programme, and why the data most factories run on doesn't describe what actually happens on the line.

Silviu learned programming by translating English menus on his father's work PC, was selling self-built accounting software at 15, and came to Germany for a PhD in machine learning. He joined VW Group Research in 2014 and spent the next eight years on perception for autonomous driving, before a question from a manufacturing colleague sent him into factories for seven months.

Today, Almetra (formerly Deltia) combines computer vision, sensors and machine data to show how production actually runs instead of how it was planned. The company raised a $19M Series A led by blisce/ in June 2026 and is opening an office in Boston.

We cover why factory data is often wrong, what happens when the MES reports a reality that isn't real, the 100% defect detection problem, why so many robotics startups get stuck in pilot purgatory, what reliability really means on a running line, and Almetra's move from measuring manual work to automating parts of it.

Silviu Homoceanu:

https://www.linkedin.com/in/silviu-homoceanu-893b7a5

Almetra:

https://www.almetra.ai/

𝐒𝐏𝐎𝐍𝐒𝐎𝐑

- CoreWeave — The Essential Cloud for AI.

CoreWeave sponsors this podcast.

Join CoreWeave at #FullyConnected:

https://coreweave.com/fully-connected-2026/agenda

👉 Free code: FCLR26

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

00:00 Meeting at the AI Campus in Berlin

00:47 Who Silviu is and what Almetra does

02:44 Romania after 1989 and his father's first PC

06:08 A 286 at home, MS-DOS, then Linux

08:04 Running a computer business at 15

10:21 High school nights in the computer lab

14:44 The mentor who said: study abroad

16:15 Flipping burgers to see the MIT campus

18:58 Romania vs Germany: theory against practice

20:23 First job: cluster computing for bridge simulations

26:20 Why computer networks stopped being enough

28:00 The exam answer that led to a PhD offer

30:47 Percentages or factors: how he picks problems

36:20 VW Research comes knocking in 2014

39:05 The meeting with manufacturing that changed everything

39:52 Seven months in factories: the data is wrong

40:30 Merantix, meeting Max, and founding the company

42:08 The Goal, bottlenecks, and life on a shop floor

43:31 The 100% defect detection problem

46:14 When the MES reports a reality that isn't real

48:09 Automating stations and the robotics angle

49:33 99.999% and why factories say no to startups

52:47 What is actually automatable today

55:07 The Series A, blisce/, and the move to Boston

59:21 Closing

Follow 22Astronauts:

Newsletter: 22astronauts.com

LinkedIn: https://www.linkedin.com/company/22astronauts

Instagram: https://www.instagram.com/_22astronauts_/

Almond is building robots for the era of physical AI.

I sat down with Saba Khalilnaji, Founder & CEO of Almond, inside the company's San Francisco factory to talk about building robots, scaling hardware, factory automation, and why Almond wants to become "America's robot factory."

Saba started building robots as a kid, later spent nearly seven years at DoorDash, and eventually left to work on one of the hardest problems in automation: getting robots to operate reliably in factories where the environment and tasks constantly change.

Today, Almond is building Axol, a dual-arm robot designed for physical AI, data collection, teleoperation and real-world deployment.

We cover why traditional automation struggles in high-mix manufacturing, what Saba learned from DoorDash, the rise of "neo-integrators," why deployments are Almond's North Star, manufacturing robots in San Francisco, and Almond's work on teaching Axol to help build Axol.

Saba Khalilnaji:

https://www.linkedin.com/in/saba-khalilnaji/

Almond:

https://www.almond.bot/

𝐒𝐏𝐎𝐍𝐒𝐎𝐑

• CoreWeave — The Essential Cloud for AI.

Join CoreWeave at Fully Connected in San Francisco, bringing together the people building the next generation of AI infrastructure and applications.

Learn more: https://www.coreweave.com/fully-connected-2026

𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒

00:00 Inside Almond's San Francisco robot factory

01:02 A first look at Axol

02:12 Building robots as a kid

07:50 Science Olympiad and early robotics

11:56 Studying bioengineering at Berkeley

14:58 From robotics to DoorDash

18:19 What DoorDash taught Saba about moving fast

22:37 Why he left to start a company

25:10 How your environment shapes ambition

29:28 Discovering the automation problem inside factories

31:06 Why traditional factory automation struggles

34:49 What Almond is building

36:03 The rise of the "neo-integrator"

38:36 Why deployments are Almond's North Star

40:17 Building America's robot factory

43:09 Why Almond believes in owning the full stack

44:05 Teaching Axol to help build Axol

46:09 Saba's advice to his younger self

47:25 Why it's called Almond

49:25 Closing

Follow 22Astronauts:

Newsletter:

22astronauts.com

LinkedIn:

https://www.linkedin.com/company/22astronauts

Instagram:

https://www.instagram.com/_22astronauts_/

Prof. Mario Krenn is the Scientific Director & Co-Founder of Feyer GmbH and Professor for Machine Learning in Science at the University of Tübingen, where he leads the Artificial Scientist Lab.

For over a decade, Mario’s research has focused on using AI not just to analyze physics data, but at a conceptual level; building AI explorers that discover completely novel experimental setups, optics, and quantum hardware. In July 2026, alongside CEO Jonathan Klimesch and CTO Sören Arlt, he co-founded Feyer GmbH: a frontier AI lab for automated physical invention.

As one of only 10 European teams selected for SPRIND's €125M Next Frontier AI Challenge (€3M non-dilutive seed phase), Feyer couples neural explorers with ultra-fast, differentiable physics simulators. Their mission is to autonomously invent next-generation industrial hardware (from advanced microscopy and lasers to lithography and quantum sensors) scaling Europe's high-tech manufacturing strengths.

Feyer GmbH: https://feyer.ai

Mario Krenn: https://mariokrenn.wordpress.com

Ilir on X: https://x.com/IlirAliu_

Ilir on LinkedIn: https://www.linkedin.com/in/ilir-aliu/

Timestamps:

0:00 "AI that invents industrial hardware": What is Feyer GmbH?

3:08 How the Cyber Valley ecosystem turned an academic researcher into a founder

5:43 Growing up in Austria: Programming operating systems at age 10

11:31 Stephen Hawking, mysteries of space, and choosing physics over engineering

15:01 Joining Anton Zeilinger's lab in Vienna: "What is the next to the next step?"

18:33 March 2014: Writing a program to design a quantum experiment humans couldn't solve

21:58 Finding solution.txt & the realization that computers can be creatively scientific

28:28 Moving to Toronto & Vector Institute: Learning AI material discovery from Alán Aspuru-Guzik

34:49 Returning to Europe & building the Artificial Scientist Lab in Erlangen & Tübingen

37:47 Asking questions to the universe: Why physics experiments are fundamental

43:35 Breaking resolution limits in microscopy & applying AI to industrial high-tech

45:49 Winning SPRIND's Next Frontier AI Challenge & founding Feyer GmbH

47:25 Scaling European hardware strengths: Semiconductor, photonics, and optics leadership.

Johannes Haux is the Co-Founder & CEO of kausable (Heidelberg, Germany), a deep-tech AI startup building reasoning-first causal foundation models.

Instead of memorizing internet-scale text patterns or relying on millions of trial-and-error attempts, kausable’s models (based on the Prior-data Fitted Network / PFN approach) learn abstract cause-and-effect structures from synthetic data—enabling them to adapt to entirely new physical environments in-context from just a handful of examples.

After working in computer vision under Prof. Björn Ommer at Heidelberg University (alongside the future Black Forest Labs founders) and serving as Head of AI at AskUI, Johannes co-founded kausable with Dr. Benjamin Herdeanu (CTO) and Gregor Ramien (COO). The company recently raised a €12M seed round led by UVC Partners and Entourage, backed by top-tier angels from OpenAI, DeepMind, Black Forest Labs, and Neura Robotics.

kausable: https://kausable.ai

Johannes on LinkedIn: https://www.linkedin.com/in/jhaux/

Ilir on X: https://x.com/IlirAliu_

Ilir on LinkedIn: https://www.linkedin.com/in/ilir-aliu/

Timestamps:

0:00 Building foundation models in the Cyber Valley deep-tech ecosystem

1:25 Working in Prof. Ommer’s lab & meeting the future Black Forest Labs team

3:48 Early startup lessons at Sysmagine & AskUI

4:09 Why internet-scale LLMs aren't enough for general-purpose physical intelligence

4:52 Discovering Prior-data Fitted Networks (PFNs) & founding kausable

7:00 "I was a bad student": Why structure and habits matter more than raw talent

12:19 Wanting to be a movie director before choosing physics at Heidelberg

18:37 Quitting academia right before COVID lockdowns & taking the entrepreneurial leap

27:38 Why community & ecosystem matter for European deep-tech startups

29:26 How humans actually learn vs. brute-force reinforcement learning in robotics

32:01 In-context learning: Adapting to distribution drift & new sensors without retraining

33:47 TipPFN & predicting critical transitions: Seizures, blackout risks, and physical dynamics

37:37 Michael Black's perspective: Why Europe is a strong launchpad for disruptive AI

41:37 Exploratory tech vs. concrete customer problems: Making the "faster horses" bet

47:17 Moving from a research lab to early design partners

50:36 "Don't fake it, but think big": Advice for European deep-tech founders

Nic Radford is the Co-Founder & CEO of Persona AI (Houston, TX) and former Founder/CEO of Nauticus Robotics (NASDAQ: KITT) & Lead of NASA’s Robonaut 2 and Valkyrie humanoid programs.

With nearly 30 years of experience building humanoid robotics (from human-rated space manipulation at NASA Johnson Space Center to subsea autonomous robots) Nic brings a veteran, unfiltered perspective to the commercial humanoid market.

Together with Co-Founders Dr. Jerry Pratt (CTO, ex-IHMC & Figure AI) and Jide Akinyode (COO, ex-NASA & Nauticus), Persona AI is building modular industrial humanoids for heavy, labor-constrained industries. Rather than chasing general-purpose home helpers or $15/hr warehouse sorting, Persona targets skilled industrial trades—starting with autonomous welding humanoids for shipbuilding in partnership with HD Hyundai and POSCO.

Persona AI: https://persona.ai/

Nic on LinkedIn: https://www.linkedin.com/in/nicolaus-radford/

Ilir on X: https://x.com/IlirAliu_

Ilir on LinkedIn: https://www.linkedin.com/in/ilir-aliu/

Timestamps:

0:00 Why the commercial humanoid robotics market has it completely wrong

3:17 Growing up in rural Indiana & fixing mechanics out of necessity

5:31 Decathlon, severe arthritis & learning "tenacity squared" as a founder

8:16 CB radios, 286/386 PCs, and entering the early Internet at Purdue

12:35 The 2000 Wired cover that sparked a 30-year humanoid obsession

15:23 Getting lost at NASA JSC and stumbling upon Robonaut 1

23:18 Camping in the Texas desert: Building amateur rockets during the early SpaceX era

25:25 Why networking is the lifeblood of deep tech partnerships

28:00 Storytelling for VCs: Why presentation skills are learned, not innate

30:50 From 15 years at NASA (Robonaut 2 & Valkyrie) to becoming an entrepreneur

35:47 Taking space tech underwater: Building Aquanaut & taking Nauticus public on NASDAQ

36:41 Partnering with Jerry Pratt to start Persona AI

37:36 Why general-purpose "Me-Too" robots face an ROI dead end

38:37 Targeting skilled trades: Why shipbuilding and heavy industry need humanoid welders

41:46 GM assembly lines & the brutal reality of automotive cycle times

44:42 The home humanoid myth: "Can we all just stop this? There is zero ROI."

48:34 First-time founders vs. Veteran founders: Tech obsession vs. Cash flow & GTM

49:50 Raising a $42M pre-seed & closing a $100M+ seed round

51:20 Delivering humanoid welders to HD Hyundai in 24 months

52:35 "Look what we built in 90 days": Calling out fake startup timelines

Prof. Dr. Jan Peters is Full Professor of Intelligent Autonomous Systems at TU Darmstadt, Department Head of Systems AI for Robot Learning (SAIROL) at DFKI, and a Founding Research Faculty Member of hessian.AI.

With over 54,000 Google Scholar citations, Jan is one of the world’s most influential robot-learning pioneers; having co-developed foundational algorithms like Natural Actor-Critic, REPS, and Probabilistic Movement Primitives (ProMPs).

IAS Lab: https://ias.informatik.tu-darmstadt.de

Jan's Website: https://www.jan-peters.net

Ilir on X: https://x.com/IlirAliu_

Ilir on LinkedIn: https://www.linkedin.com/in/ilir-aliu/

Timestamps:

0:00 Combining machine learning with physical robotics before it was cool

4:05 Studying 4 Master’s degrees & dodging the German military draft

8:27 Why DLR and ATR in Japan were lightyears ahead of German universities

12:37 Meeting Kawato, Chris Atkeson, and Stefan Schaal in a bamboo forest

17:10 Turning down Stanford and CMU for a smaller lab at USC

21:46 How double-blind reviews leveled the academic playing field

23:44 Injecting physical mechanics into reinforcement learning

26:01 How 9/11 permanently shifted American culture and optimism

29:40 Returning to Max Planck: The "monastery" era of AI research

35:14 Why being authentic & speaking your mind wins long-term respect

44:01 Surviving hostile German hiring commissions in the 2000s

47:55 Why TU Darmstadt became Europe’s secret powerhouse for physical AI

57:39 High-speed table tennis, juggling, and the limits of robot perception

58:32 Raising €2M for humanoids and fighting 4 years of European bureaucracy

65:49 German red tape vs. American risk-taking: What's holding Europe back?

68:33 Why TU Darmstadt rivals top US universities in AI breadth