Audrow Nash Podcast

Episode 27 of the Audrow Nash Podcast.

A four-and-a-half-foot humanoid kicks in a door so a person doesn't have to go in first. Amarjot's case is that work like this is where humanoids pay off now: the tasks vary less than in a home, and working 70 or 80% of the time is still worth it when a person isn't the one getting shot.

I talk with Amarjot Singh, founder and CEO of Skylark Labs, about a model that flags, labels, and learns new things on the device after it's deployed, how they teach a humanoid a skill from a few demonstrations, and why he thinks humanoid hardware becomes a commodity while the value moves to the brain.

You'll like this interview if you work on robot learning or perception at the edge, you follow humanoids, or you're curious how robots get sold into defense. The model architecture stretch (8:51–34:41) is the most technical.

EPISODE LINKS
- Skylark Labs website: https://www.skylarklabs.ai/

PODCAST LINKS
- Podcast Website: https://audrownashpodcast.com/
- YouTube: https://www.youtube.com/@audrow
- Apple Podcasts: https://podcasts.apple.com/us/podcast/audrow-nash-podcast/id1716486786
- Spotify: https://open.spotify.com/show/74jWpWiLwsasY2QHtDcl8I?si=6c92796bf9554162

TIMECODES
0:00 Cold open
0:33 Amarjot Singh and Skylark Labs
3:30 Humanoids that sweep for bombs and breach doors
7:23 A brain-inspired model with memory
8:51 Wavelets and Gabor filters as the first layer
13:40 Unsupervised features, supervised on top
17:26 1,000x less data and flagging outliers
21:58 Detecting a camouflaged tank
25:53 Explainability for kinetic systems
28:12 Self-labeling and on-device memory
32:06 From perception to motion
34:41 The robotics stack and sensor language model
36:45 Skills stored as graphs
39:10 Mapping human motion onto a humanoid
40:33 Adapting to a new door handle
43:18 Teaching with Apple Vision Pro and egocentric video
46:58 Simulation and sim-to-real
50:14 What's easy, hard, and not yet possible
52:44 When 70% reliable is enough
55:01 The humanoid Turing test
56:51 Generalization is the bottleneck
1:00:05 From runway debris to robot dogs
1:03:46 Why a humanoid, on two legs or four
1:07:00 Humanoid hardware as a commodity
1:10:56 Why defense comes before homes
1:14:32 Trying markets until someone paid
1:18:40 Building their own humanoid
1:19:48 Manufacturing humanoids in the US
1:22:20 India for data, testing, and talent
1:25:40 Skylark's next three to five years
1:28:21 Advice for robotics founders
1:30:02 Where robotics is going
1:31:57 Final takeaway

Episode 26 of the Audrow Nash Podcast.

Johns Hopkins surgeons showed Axel's team the clip-and-cut phase of gallbladder surgery. The team recorded demonstrations, trained a robot on them, and it did that phase eight times in a row with no human help. The organs were from a butcher. Now they're testing it in live animals.

I talk with Axel Krieger, Chief Robotics Officer at Inner Logic and associate professor at Johns Hopkins, about why his lab moved from hand-built models to imitation learning from surgeons, how they break a procedure into small steps a robot can learn, and how simulated patients built from tens of thousands of CT scans help test surgical devices.

You'll like this interview if you're a roboticist curious about surgery, you work on imitation learning or VLAs, or you want a grounded take on when autonomy reaches the operating room. The high- and low-level policy stretch (23:46–42:28) is the most technical.

EPISODE LINKS
- Axel's LinkedIn: https://www.linkedin.com/in/axel-krieger-b14a0075/
- IMERSE Lab (Johns Hopkins): https://imerse.lcsr.jhu.edu/
- Inner Logic website: https://innerlogic.io/

PODCAST LINKS
- Podcast Website: https://audrownashpodcast.com/
- YouTube: https://www.youtube.com/@audrow
- Apple Podcasts: https://podcasts.apple.com/us/podcast/audrow-nash-podcast/id1716486786
- Spotify: https://open.spotify.com/show/74jWpWiLwsasY2QHtDcl8I?si=6c92796bf9554162

TIMECODES
0:00 Cold open
0:39 Axel Krieger and Inner Logic
4:05 How surgical robots work today
9:41 AI infrastructure for device makers
11:15 From stitching to imitation learning
18:23 Autonomous gallbladder surgery steps
21:37 Action chunking transformers (ACT)
23:46 Breaking surgery into small steps
29:42 Language between high and low level policies
35:44 Adapting to surgical robot hardware
40:01 How much training data it takes
42:28 VLAs and large surgical datasets
47:30 Mixing training data
51:44 From dead tissue to live animals
54:47 Self-driving analogy and FDA approval
58:35 Inner Logic's data platforms
1:02:42 Simulated patients from CT scans
1:05:20 Simulating bones, vessels, and soft tissue
1:13:07 Where surgical robotics is heading
1:17:28 One surgeon, several surgeries
1:18:57 Training the next surgeons
1:22:19 Funding climate
1:24:52 Next milestones, from suction to suturing
1:26:58 Tumor resection and full procedures
1:29:49 Robot surgery in 20 years?
1:31:15 Advice for getting into surgical robotics
1:35:02 Final takeaway

Wolf's goal for Pixi: git clone a repo, then something like pixi run start sets up dependencies and actually runs the project. Cross-platform. Not "hope Ubuntu's apt has the right CMake."

I talk with Wolf Vollprecht (founder & CEO) and Ruben Arts (software engineer) at Prefix.dev about Pixi as a Conda-ecosystem package manager rewritten in Rust, RoboStack turning ROS packages into Conda binaries (Linux, Mac, Windows subsets), and why robotics teams get stuck juggling apt, pip, and the ROS index.

You'll like this if you ship ROS workspaces, teach workshops, or care about Mac/Windows robotics tooling beyond "build from source and pray." The stack explainer (about 4:32–19:30) is the densest conceptual stretch; demos start around 25:50.

EPISODE LINKS

- Prefix.dev: https://prefix.dev/

- RoboStack: https://robostack.github.io/index.html

- Wolf Vollprecht LinkedIn: https://www.linkedin.com/in/wolfvollprecht/

- Ruben Arts LinkedIn: https://www.linkedin.com/in/ruben-arts/

- Wolf on X: https://x.com/wolfvollprecht

- Ruben on X: https://x.com/rarts_ruby

PODCAST LINKS

- Podcast Website: https://audrownashpodcast.com/

- YouTube: https://www.youtube.com/@audrow

- Apple Podcasts: https://podcasts.apple.com/us/podcast/audrow-nash-podcast/id1716486786

- Spotify: https://open.spotify.com/show/74jWpWiLwsasY2QHtDcl8I?si=6c92796bf9554162

TIMECODES

0:00 Cold open

0:50 Guest intros

1:44 Motivation behind Pixi

4:32 Package management basics

9:24 Robotics setup pain

14:43 RoboStack cross-platform

19:30 Prefix business model

25:50 Demo: setting up Pixi

46:49 Quick setup and caching

49:00 Building ROS packages

51:01 Features and activation

52:55 Tasks in Pixi

59:37 Building Conda packages

1:16:01 Future and community

Episode 25 of the Audrow Nash Podcast.

A bag of gene-edited cells can be worth about half a million dollars (Fred's range for a dose is roughly $400K to $2M), and today that work is still done by hand in clean rooms, over thousands of steps.

I talk with Fred Parietti, Co-founder and CEO of Multiply Labs, about robots that manufacture cell and gene therapies using the same instruments, reagents, and consumables the scientists use, why they train from video of scientists instead of teleoperation, and where a ~$30K humanoid fits in their plant (loading and unloading carts, not the science).

You'll like this interview if you're a roboticist curious about pharma, a founder picking high-value problems over pizza robots, or anyone who wants Fred's take on "infinite market" pitches and a robotics winter. The biology explainer (16:08–34:00) is the most technical stretch.

EPISODE LINKS

- Multiply Labs: https://www.multiplylabs.com

PODCAST LINKS

- Podcast Website: https://audrownashpodcast.com/

- YouTube: https://www.youtube.com/@audrow

- Apple Podcasts: https://podcasts.apple.com/us/podcast/audrow-nash-podcast/id1716486786

- Spotify: https://open.spotify.com/show/74jWpWiLwsasY2QHtDcl8I?si=6c92796bf9554162

TIMECODES

0:00 Cold open

0:53 Multiply Labs overview

7:17 Pharma robotics gap

16:08 Next-gen pharma

34:00 Multiply Labs automation

43:54 Deployment size

1:00:04 Thoughts on humanoids

1:10:01 Training process

1:21:39 High-value problems

1:24:37 Key takeaways

Episode 24 of the Audrow Nash Podcast.

For a wheeled robot, "shut it off" can be a safety move. For a biped that has to stay upright, killing power is the hazard. Pras walks through why Digit needs different rules.

I talk with Pras Velagapudi, CTO of Agility Robotics, about Digit (two arms, two legs) doing paid material handling on full shifts, why a dynamically stable biped beats a heavy wheeled base for narrow aisles, the tradeoffs of wheels on feet, and how they're working with others (including Boston Dynamics) on safety standards for robots that can't just lock their wheels and call it safe.

You'll like this interview if you're into humanoids, warehouse automation, or the control and safety problems that show up once a robot has to balance. The legs-and-wheels stretch (about 5:26–23:15) is the densest design argument; safety and standards (36:42 on) is where the shut-it-off problem shows up.

EPISODE LINKS

- Agility Robotics: https://www.agilityrobotics.com

PODCAST LINKS

- Podcast Website: https://audrownashpodcast.com/

- YouTube: https://www.youtube.com/@audrow

- Apple Podcasts: https://podcasts.apple.com/us/podcast/audrow-nash-podcast/id1716486786

- Spotify: https://open.spotify.com/show/74jWpWiLwsasY2QHtDcl8I?si=6c92796bf9554162

TIMECODES

0:00 Cold open

0:42 Agility overview

5:26 Why humanoid design

13:26 Why legs, not wheels

23:15 Digit's application

36:42 Safety in humanoids

49:11 Standardizing safety

1:06:50 Digit in other industries

1:14:41 How AI fits into robotics

1:29:23 Is data still gold?

1:35:54 The future of Agility

1:38:13 Key takeaways

Episode 23 of the Audrow Nash Podcast.

Stefan's cold open: the best robot in the world today is less valuable to its owner than an average computer was in 1960. They argue robotics still makes teams rebuild the same navigation stack instead of buying commodity software.

I talk with Stefan Seltz-Axmacher (CEO) and Ilia Baranov (CTO), co-founders of Polymath Robotics, about packaging point-to-point navigation for off-highway machines (tractors, bulldozers, mines), why SaaS feels easy because you don't rebuild payments and backends, labor shortages and vehicle SKU sprawl, and why they think humanoid-as-general-purpose is often the wrong bet next to Roomba-scale appliances.

You'll like this if you build autonomy for industrial vehicles, sell into mining/ag, or want a blunt take on demos vs products. The ROI/history stretch (about 12:14–24:15) and the humanoid/minimally-viable-humanoid bit (52:02 on) are the densest arguments.

EPISODE LINKS

- Polymath Robotics: https://www.polymathrobotics.com

PODCAST LINKS

- Podcast Website: https://audrownashpodcast.com/

- YouTube: https://www.youtube.com/@audrow

- Apple Podcasts: https://podcasts.apple.com/us/podcast/audrow-nash-podcast/id1716486786

- Spotify: https://open.spotify.com/show/74jWpWiLwsasY2QHtDcl8I?si=6c92796bf9554162

TIMECODES

0:00 Cold open

0:12 Polymath overview

6:04 Robotics standards gap

12:14 Market reality

16:43 Labor shortages

24:15 Cutting mining costs

33:55 AgTech needs action

43:16 Practical autonomy

52:02 Humanoid future

1:12:31 LLMs and coding

1:21:10 Work-kids balance

1:36:04 Key takeaways

Episode 22 of the Audrow Nash Podcast.

Twenty years ago, SaaS had no rule of 40. Brannon Jones's take is that deep tech is in that same stretch now: the rules are still unwritten.

I talk with Brannon Jones, investor on AlleyCorp's deep tech team, about why they started the vertical in robotics, how they underwrite hardware with a value-creation multiple instead of SaaS metrics, and what that looks like in the portfolio: EyeBot's 90-second eye exam kiosk, Portal Space's satellite thrusters, and Aescape's robotic massage table (which he says raised $83M). We also get into why AlleyCorp hasn't backed a humanoid yet, and why he thinks a lot of embodied-AI demos are still pick and place.

You'll like this if you raise for a robotics company, invest in hardware, or want an operator-turned-investor view of New York deep tech. The portfolio walk (about 15:43–56:00) is concrete; the investing framework (56:08 on) is the densest stretch.

EPISODE LINKS

- Brannon Jones X: https://x.com/robotjonez

- Brannon Jones LinkedIn: https://www.linkedin.com/in/jones-brannon

- AlleyCorp website: https://alleycorp.com

- AlleyCorp LinkedIn: https://www.linkedin.com/company/alleycorp

PODCAST LINKS

- Podcast Website: https://audrownashpodcast.com/

- YouTube: https://www.youtube.com/@audrow

- Apple Podcasts: https://podcasts.apple.com/us/podcast/audrow-nash-podcast/id1716486786

- Spotify: https://open.spotify.com/show/74jWpWiLwsasY2QHtDcl8I?si=6c92796bf9554162

TIMECODES

0:00 Cold open

0:23 AlleyCorp and Brannon Jones

3:26 What deep tech means at AlleyCorp

10:14 New York vs West Coast portfolio

15:43 Deep tech portfolio overview

17:11 EyeBot eye exam kiosk

22:25 Why deep tech vs SaaS

28:46 Portal Space satellite thrusters

32:35 Aescape massage robot

44:39 Valar Atomics

49:24 Koop Technologies

50:58 Glacier recycling robots

54:46 Civ Robotics solar surveying

56:08 Funding robotics without SaaS playbooks

1:09:50 Market-focused startups

1:14:17 Early robotics investments

1:23:40 Humanoids and value creation

1:28:32 Deglobalization and China

1:32:48 Hype vs pick and place

1:38:19 Key takeaways

Episode 21 of the Audrow Nash Podcast.

Somebody draws a house on a napkin. BotBuilt turns that into a 3D nail schedule, cuts twisted lumber in a warehouse, and ships panels so on-site framing drops from weeks to about four hours.

I talk with Barrett Ames, co-founder of BotBuilt, about why they automate framing (the bones of the house) instead of the whole build, how they plan and grip imperfect lumber, and the business of warehouse construction that still has to win on jobsites.

You'll like this if you care about construction robotics, founder-market fit in hardtech, or how to pick an automation wedge that isn't a race to the bottom. The lumber-handling stretch (about 6:59–17:41) is the densest engineering bit; time and cost (56:52 on) is the densest business bit.

EPISODE LINKS

- BotBuilt: https://www.botbuilt.com

- Barrett Ames: https://www.cbames.com

PODCAST LINKS

- Podcast Website: https://audrownashpodcast.com/

- YouTube: https://www.youtube.com/@audrow

- Apple Podcasts: https://podcasts.apple.com/us/podcast/audrow-nash-podcast/id1716486786

- Spotify: https://open.spotify.com/show/74jWpWiLwsasY2QHtDcl8I?si=6c92796bf9554162

TIMECODES

0:00 Cold open

2:22 Overview

6:59 Challenges

17:41 Sketch to structure

27:20 Generative solutions

30:36 Structuring build

35:49 When things go wrong

42:31 Traditional vs automated

56:52 Time and cost

1:07:36 Long-term vision

1:14:35 Construction opportunities

1:23:51 Lessons learned

1:33:32 Final thoughts

Episode 20 of the Audrow Nash Podcast.

Tesla gets far on vision alone, but shipping that stack means hardware, bandwidth, AI, and sync. Luxonis's bet is cameras that are eyes, ears, and a chunk of brain at the edge.

I talk with Bradley Dillon, CEO of Luxonis, about AI-enabled cameras for robots (he cites more than 150,000 sold), fusing sensors on-device, multi-device linking, models and labeling, and where they aim the platform next (including harder industrial and security screening angles).

You'll like this if you build robots and don't want to become a camera company, or if you care about edge vs cloud for vision. The edge-compute stretch (about 25:01–31:41) and applications (20:26 on) are the densest product bits.

EPISODE LINKS

- Luxonis: https://www.luxonis.com

- Bradley Dillon LinkedIn: https://www.linkedin.com/in/bradley-dillon/

PODCAST LINKS

- Podcast Website: https://audrownashpodcast.com/

- YouTube: https://www.youtube.com/@audrow

- Apple Podcasts: https://podcasts.apple.com/us/podcast/audrow-nash-podcast/id1716486786

- Spotify: https://open.spotify.com/show/74jWpWiLwsasY2QHtDcl8I?si=6c92796bf9554162

TIMECODES

0:00 Cold open

1:54 Bradley Dillon and Luxonis overview

8:22 Luxonis equipment

13:01 Integrating multiple sensors

20:26 Applications of Luxonis vision

25:01 Luxonis and computing power

28:21 Linking multiple devices

31:41 Luxonis AI models

34:53 Data generation and labeling

44:53 Future predictions

49:01 AI for scene understanding

59:31 Targeting industries

1:04:54 Smart security screening

1:09:43 Growth story

1:18:58 Chip supply challenges

Episode 19 of the Audrow Nash Podcast.

Three short live interviews from a stealth robotics meetup between Austin and San Antonio: legged-robot safety in human spaces, automating mining exploration tasks, and robots that frame houses.

I talk with Kyle Morgenstein (UT Austin PhD, legged safety / RL), David Venegas (Head of Operations at Durin, mining), and Barrett Ames (co-founder of BotBuilt, construction framing).

You'll like this if you want a sampler across research and startups rather than one deep company dive. Kyle's safety stretch opens (about 1:24–20:36); David covers mining (20:36–38:03); Barrett closes on BotBuilt (38:03 on).

EPISODE LINKS

- Kyle Morgenstein: https://www.kylemorgenstein.com/

- Durin: https://www.durin.com/

- BotBuilt: https://www.botbuilt.com/

PODCAST LINKS

- Podcast Website: https://audrownashpodcast.com/

- YouTube: https://www.youtube.com/@audrow

- Apple Podcasts: https://podcasts.apple.com/us/podcast/audrow-nash-podcast/id1716486786

- Spotify: https://open.spotify.com/show/74jWpWiLwsasY2QHtDcl8I?si=6c92796bf9554162

TIMECODES

0:00 Cold open

1:24 Kyle: legged robot safety

3:55 Improving robot safety for humans

7:25 Humanoid progress and future

10:39 Boston Dynamics vs Unitree

12:14 Reward models for safer robots

15:05 Foundation models

16:34 Teleoperation efficiency

18:45 Training from YouTube data

20:36 David: mining basics

28:58 Automating mining tasks

34:09 Access challenges in excavation

36:57 Use of simulation

38:03 Barrett: construction robots

41:27 Custom home focus

43:06 BotBuilt's approach

44:16 BotBuilt's data sourcing

49:47 Working with robots

49:57 Perception stack

52:20 Handling lumber challenges

53:39 Margins vs prefab homes

53:57 Transport costs

56:02 BotBuilt's lifecycle

58:17 Outro