Deepak Pathak & Abhinav Gupta (S... | NVIDIA AI Podcast Summary

NVIDIA AI Podcast | Deepak Pathak & Abhinav Gupta

1. Key Themes

The "Horizontal Platform" Thesis for Robotics

Skild's core bet is that robotics is about to undergo the same transformation language did when LLMs replaced verticalized chatbots and search engines. Rather than building task-specific robots, they are building a shared foundational brain that gets fine-tuned per deployment — the same way GPT is fine-tuned for specific enterprise applications.

"Language also before this whole came in was very verticalized. There were some different companies building chatbots. There were different companies building search engines. But once LLM came in, they became the horizontal platform. And now everyone is building on top of that horizontal LLM platform. That is exactly how we are now thinking about robotics." — Abhinav Gupta [00:04:40]

Robotics Is Fundamentally a Data Problem, Not a Hardware Problem

The founders repeatedly frame the core bottleneck in robotics as data scarcity, not compute or hardware. This reframing has major implications for how one evaluates any robotics company — the competitive moat is in data architecture and accumulation, not mechanical engineering.

"Robotics is a data problem. Unlike language or vision, there is not much data in robotics. There is no internet of robot data. So if that's the scenario, we cannot pick and choose which data we use." — Deepak Pathak [00:00:00]

Three-Source Data Architecture: Pre-Train, Simulate, Post-Train

Skild uses a deliberate three-layer data strategy that mirrors how LLMs are built: video data for broad pre-training, simulation for robustification at scale, and small-batch real-world robot data for precision post-training. Each layer addresses the weaknesses of the others.

"We use the video data to pre-train our models... However, the problem with videos is if we can learn everything from videos, Deepak gives this great example that if we can learn from videos, all of us would be Federers... So that is where for us simulation comes into play... But again, sim to real gap still exists. And now we take this model which has been pre-trained on videos and simulation. But before deployment, we post train it on the real world data." — Abhinav Gupta [00:11:08]

The Self-Sustaining Data Flywheel as a Competitive Moat

Every deployment feeds data back into the shared brain. Structured environments (factories, warehouses) serve as the first flywheel stage, generating data that makes the system capable of handling less structured environments (hospitals, hotels), which eventually bootstraps consumer/home robots. This sequencing is the company's deliberate go-to-market architecture.

"You start with factories. They act as data flywheel for semi-structured scenarios like hospitals, grocery stores, hotels. The flywheel from there helps you get to the ultimate challenge, which is like homes, consumer robots. So this is basically how we are orchestrating the self-sustaining data flywheel loop from every deployment." — Deepak Pathak [00:16:36]

Deployment Is a Technical Challenge, Not an Afterthought

Unlike software AI products where deployment is near-instant once a model is ready, physical AI deployment is itself a hard engineering problem. Skild treats it as a first-priority technical challenge from day one, which differentiates them from research-first robotics organizations.

"For us, deployment is our first priority from day one... Deployment is in itself a big technical challenge. And how do you orchestrate that at scale? It has not been done before." — Deepak Pathak [00:02:47] and [00:29:06]

The 90% Wall: Why Traditional Robotics Never Went Mainstream

Traditional robotics could reach 80–90% automation performance through precise programming, but the long tail of corner cases in the physical world made full automation impossible — keeping humans perpetually in the loop. AI-driven generalist models solve this by treating one vertical's corner cases as another vertical's training data.

"It's very easy to get the first 80% or 90% of the performance. But then you hit this wall, which is called the corner cases in the physical world... And that is why it has not been — traditionally robotics has not really gone big mainstream essentially." — Abhinav Gupta [00:03:45]

Omnibodied Intelligence: Form-Factor Agnosticism as a Strategic Choice

The same brain controlling a humanoid, a dog-like robot, and a factory arm simultaneously is not just a technical feat — it is a deliberate strategy to maximize data diversity. Every form factor adds a different class of physical interaction data, making the brain more robust.

"You can have a humanoid, or a dog-like robot, or a robotic arm on a conveyor belt, all being controlled by the same shared brain, shared intelligence behind the scene." — Deepak Pathak [00:00:56]

Edge Compute Is the Next Hardware Frontier for Physical AI

The GPU-in-a-server paradigm that powers LLMs does not work for robots that must react in real time without network round-trips. On-device edge inference is a fundamental hardware requirement for physical AI — and a major area of co-development with NVIDIA.

"The solution that worked for LLMs of big GPUs in servers, it will look very different for a robot. Because robot doesn't have time to connect to a server if it's falling. React immediately. So on device edge compute, this is where we are partnering as well." — Deepak Pathak [00:19:04]

2. Contrarian Perspectives

The "Fastest Growing Product" Playbook Does Not Apply to Physical AI

Most tech investors and operators assume AI products can achieve viral, near-instant adoption once ready. The founders explicitly reject this for robotics — physical deployment is inherently slow, iterative, and cannot be compressed by product marketing.

"Physical AI is not like that. The things take time to deploy. So for us, deployment is our first priority from day one... For instance, in case of ChatGPT or language models, folks did research for several years. But once it was ready, you have a hundred million users in one month. Right. Fastest growing product." — Deepak Pathak [00:02:47]

Home Robots Are Not the Starting Market — They Are the End State

Conventional excitement around consumer humanoids and home robots misunderstands the deployment sequencing. The founders argue that factories and warehouses must come first — not because they are more interesting, but because they are the necessary data substrate for everything that follows.

"In this year itself, we will start to see deployments in factory warehouse around people that bootstraps the next one, like hospitals, hotels, service industry — that bootstraps the ultimate consumer robots. It's very hard to put the timeline for the ultimate home robots." — Deepak Pathak [00:24:40]

Watching Video Is Not Sufficient to Learn Physical Skills — At Any Scale

There is an implicit assumption in some AI research circles that sufficiently large video datasets could teach robots physical manipulation. The founders reject this, using an intuitive human analogy that watching Federer play tennis has never made anyone play like Federer.

"If we can learn from videos, all of us would be Federers. Because we will watch Federer and we'll start playing like Federer and so on. So that's never going to be sufficient. Just watching videos is not going to be sufficient. If it was sufficient, I could dunk a basketball, but I can't." — Abhinav Gupta [00:11:08]

Experts Are Systematically Underestimating Long-Term AI Progress

The founders — with 20+ years in the field each — admit that even they have been consistently surprised by the pace of progress, and are now reluctant to make public predictions. This is a signal that published expert forecasts about robotics timelines are likely conservative.

"The progress of compute and the hardware coming costs coming down has just made this all so surprising that I would say even the experts like us who have been working in this for 20 years are scared to say anything online." — Abhinav Gupta [00:26:28]

Hardware Reliability, Not AI Capability, Is the Real Bottleneck for Home Robots

The common framing is that AI intelligence is what's missing for home robots. The founders point to a less-discussed constraint: current humanoid hardware is simply not reliable or safe enough to be deployed in homes, regardless of how good the AI gets.

"While you're seeing so much hardware in humanoid space also, are these hardware reliable to be even put in homes today? Like no one has put them because safety again is a big issue. Like when you are putting them in home, what if it falls and there's a child around?" — Abhinav Gupta [00:25:33]

3. Companies Identified

Skild AI Pittsburgh-based robotics AI company building "OmniBrain," a universal foundational model for robotics across all form factors and tasks. Founded by two Carnegie Mellon professors. Two and a half years old at time of recording. HQ in Pittsburgh, offices in San Mateo and Bangalore.

"At Skild, we are building a general purpose brain. So we call this OmniBodied Intelligence. Any robot, any task, one brain." — Deepak Pathak [00:00:56]

NVIDIA Semiconductor and AI platform company. Named as a deep technology partner for Skild across simulation (Isaac, Isaac Gym, Newton physics solver), video/generative models (Cosmos), and edge compute hardware.

"Our company is two and a half years old, but I have been working personally with NVIDIA since 2018... We are now working with NVIDIA on Newton as well. And in fact, we are co-developing better physics solvers. We'll probably open source them together." — Deepak Pathak [00:18:09]

4. People Identified

Deepak Pathak Co-founder of Skild AI, former Carnegie Mellon University professor. One of the leading researchers in robot learning. Has been collaborating with NVIDIA since 2018. Co-architect of the OmniBrain system and its deployment strategy.

"We both have been professors before this. So we are extremely technical. We have been involved in bringing up these technologies in the robot learning area for the last decade and more." — Deepak Pathak [00:01:49]

Abhinav Gupta Co-founder of Skild AI, former Carnegie Mellon University professor. Co-architect of Skild's data strategy and deployment pipeline. Frames the horizontal-platform analogy for robotics and articulates the three-source data architecture.

"We are almost rethinking the way robotics is done traditionally... We are building this horizontal general purpose brain that can then be fine-tuned for different verticals essentially." — Abhinav Gupta [00:03:45]

5. Operating Insights

A Three-Layer Testing Protocol for Physical AI Deployment

Before any deployment, Skild runs a structured three-stage QA pipeline: (1) task-specific KPIs (accuracy, cycle time), (2) generalization stress tests (unexpected objects, lighting changes, environmental variation), and (3) safety guardrail verification (hardware failure modes like cut camera wires, boundary enforcement). This is a replicable framework for any company deploying physical AI systems.

"Our testing has these KPIs that we first test on... But just doing KPIs is not sufficient... We go and test for generalization. We say, OK, what if someone left a box here? Or what if somehow the lights were completely off?... And last is the safety — that in no scenario should you break the safety violations." — Abhinav Gupta [00:20:02]

Deploy as a Generalist First, Specialize Through Post-Training

The operational model for scaling new robot tasks is: (1) attempt off-the-shelf deployment with the generalist brain, (2) if insufficient, collect a few days of domain-specific data or use simulation assets, (3) post-train the model on that small dataset, then deploy. This dramatically reduces the time and cost to stand up a new robotic application.

"Let's say now you go to a different task... we may collect data for a few days. OK, either do that or if you already have the assets, then we'll get it in simulation either way. Then we use the data and we post train the model and then that model takes over and it turns on the robot directly." — Deepak Pathak [00:14:49]

Start the Data Flywheel Early — It Takes Time to Build Momentum

The founders explicitly note that a data flywheel cannot be activated at the moment you need it — it requires early, broad deployment to accumulate the momentum that makes future tasks cheaper and faster to solve. Companies building physical AI should prioritize deployment breadth over perfection to start the flywheel.

"Flywheel takes time to set up, takes time to get momentum. And if these things are to happen in the timeline we want them to happen, we have to start now." — Deepak Pathak [00:28:10]

6. Overlooked Insights

Open-Source Physics Solvers Could Become a Critical Infrastructure Layer

In a single throwaway line, Deepak mentions that Skild and NVIDIA are co-developing improved physics solvers for the Newton simulator and will likely open-source them together. This is potentially a foundational infrastructure contribution — a better physics engine that reduces the sim-to-real gap would benefit the entire robotics AI ecosystem and could become the standard substrate for training physical AI models, similar to how PyTorch became the standard for deep learning.

"We are now working with NVIDIA on Newton as well. And in fact, we are co-developing better physics solvers. Probably will open source them together." — Deepak Pathak [00:18:09]

The "Generalist-to-Specialist Data Pull-Back" Loop Is a Novel AI Architecture

Briefly described and then moved past, this concept deserves attention: specialists trained from a generalist brain can feed their domain-specific data back into the shared generalist brain — something impossible in human cognition but trivially achievable in software. This creates a compounding intelligence loop that no single-vertical robotics company can replicate, and represents a durable structural advantage that will widen over time.

"When you have this specialist, then the data can pull back from all of them and come to the same brain behind the scene, which is not what happens in humans, but we can do it in a computer. And now when you are the next task to go to, you will need less data for the next task." — Deepak Pathak [00:15:36]