Apr 26, 2025 · 30m · tbpn
The Race to Create General-Purpose Robots | Karol Hausman & Lachy Groom on TBPN
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Physical Intelligence co-founders Karol Hausman and Lachy Groom discuss the development of Pi-05, an end-to-end foundation model designed to achieve zero-shot physical intuition and general-purpose robotic manipulation across diverse real-world environments.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 25.2% of the talking time here. How this is scored →
The hosts as informed peer 4.4 Guest teaching 4.2 Guest disagreement 1.2 The hosts pushing back 1.4
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 28:20 Lachy calls out deceptive teleoperated Twitter demos
Lachy Groom forcefully counters industry hype, dismissing flashy backflip videos and teleoperated demos as easy parlor tricks compared to real long-horizon research.
Hardest push from the hosts ▶ 27:56 John presses on the Waymo vs Tesla ontology for roboticsJohn Coogan refuses to let the self-driving analogy drop, directly challenging Groom on whether venture capitalists should maintain a dual Waymo-versus-Tesla framework.
Biggest teaching moment ▶ 7:50 Karol breaks down the physics simulation dichotomyKarol Hausman lucidly educates the hosts on why simulation easily solves locomotion by modeling internal kinematics, but breaks down during physical object manipulation.
The host holds their own ▶ 15:24 John breaks down deterministic control vs end-to-end AI scalingJohn Coogan articulates an informed technical overview comparing deterministic C++ control loops with data-driven end-to-end scaling flywheels.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Founding Physical Intelligence and Leaving Big Tech | 3 | 3 | 1 | 1 | Coogan and Hays ask friendly introductory questions about Physical Intelligence's founding and the new Pi-05 milestone. Lachy Groom gently reframes Hays's consumer timeline assumptions by explaining that they are operating purely as a research lab right now rather than targeting 50% reliability consumer rollouts. | |
| Adapting Multimodal Architectures to Robotic Data Scarcity | 4 | 4 | 1 | 1 | Coogan asks about foundational model turning points like transformers. Karol Hausman explains that unlike LLMs with the entire open internet, robotics requires discovering custom recipes where small percentages of diverse multi-embodiment data drive broad generalization. | |
| The Limits of Simulation in Physical Object Manipulation | 5 | 5 | 1 | 1 | Coogan astutely distinguishes between procedurally generated 2D navigation simulations and complex soft-body physics manipulation. Hausman validates the host's premise and expands on why simulation solves internal body locomotion easily while falling short on diverse object manipulation. | |
| Data Scaling Logistics and the Robotics Startup Ecosystem | 4 | 4 | 1 | 2 | Coogan pushes on the logistical feasibility of gathering real-world home data at scale. Groom and Hausman explain their multi-pronged data collection strategy and reveal that generalization requires far fewer unique physical environments than previously assumed. | |
| From Google's Arm Farm to Zero-Shot Physical Intuition | 6 | 4 | 1 | 1 | Coogan demonstrates domain knowledge by asking about Karol's co-founders and the historical Google Arm Farm. Hausman explains how the arm farm proved reinforcement learning works for grasping but notes modern models now possess zero-shot physical intuition instead of aimless flailing. | |
| The Shift to Pure End-to-End Deep Learning | 6 | 5 | 2 | 2 | Coogan demonstrates deep technical framing regarding deterministic C++ control stacks versus scaling laws in end-to-end models. Hausman immediately clarifies that end-to-end robotics is already deployed in their current models because rule-based programming fundamentally cannot handle messy physical tasks like laundry folding. | |
| High-Velocity Culture, Talent Alignment, and Hardware Supply Chains | 3 | 3 | 1 | 1 | Coogan and Hays ask about operational lessons from Stripe and Google and the impact of hardware tariffs. Groom details their 'alignment tax' hiring philosophy and explains why subscale R&D allows time to develop domestic supply chains. | |
| Compute Infrastructure and the Boundless Real-World Data Frontier | 5 | 5 | 1 | 1 | Coogan inquires whether massive gigawatt-scale data centers like Stargate are required for physical models. Hausman educates the hosts by explaining that robotics is not yet compute-bottlenecked, but robotics will eventually supply frontier LLMs with limitless real-world interactive data. | |
| Deconstructing the Self-Driving Analogy and Research Horizons | 4 | 5 | 2 | 3 | Coogan prompts the guests to deconstruct the self-driving analogy and presses on whether VCs should categorise robotics companies like Waymo vs Tesla. Groom pushes back against hype, warning that social media demos are often teleoperated tricks and framing their real competitor as the frontier of fundamental science. |
Statements from this episode (23)
Karol Hausman: Physical Intelligence is building a universal robot foundation model
“We want to build a model that can control any robot to do any task.”
Hausman: Robotics' biggest bottleneck is generalization, not dexterity
“The biggest challenge in robotics so far hasn't really been Agility or dexterity, what the robots can do. But then generalization.”
Hausman: Pi-05 model achieves 50% to 80% task success in unfamiliar homes
“And it turns out that with Pi oh five, which we just released yesterday, we can do that. And it doesn't work all the time. It's not that I can just give it to you and it will work in your kitchen every single time, but it works quite often quite well. So we br…”
Groom: Consumer robotics deployment will require 98% to 99% reliability
“I think there'll be a point at which it gets good enough that we can deploy it to consumers, but it's not going to be, like, 50%. It's going to be closer to 98, 99%.”
Hausman: Diverse data across robotic form factors and tasks cross-transfers
“And it turns out if you collect very diverse data across many different tasks from many different form factors, they all contribute to each other. And that they contribute to a better understanding for the model of what actually is happening and how to utilize…”
Hausman: In-home mobile manipulator data is a tiny training dataset fraction
“Interestingly, most of the data is actually not the model manipulators in many different homes. It's a very, very small percentage of it.”
Hausman: Simulation fails for robotic manipulation due to real-world object diversity
“It hasn't worked nearly as well for manipulating objects or working with your hands, and I think the reason for that is then the difficulty isn't about, like, how do you move your hands? It's more about the world that you're manipulating, and that is much hard…”
Hausman: Robot models require surprisingly few environments to generalize to new ones
“So far we've been quite surprised by how few different environments you need to see to be able to generalize to a new one.”
Groom: Physical Intelligence works with most new robotics startups globally
“We work with, I'd say probably most of the new robotics companies starting in the U S and abroad.”
Hausman: Physical Intelligence models give robots zero-shot intuition in new environments
“And one thing with Pi-O-V that we are really excited about is that we are now at the stage where the robots kind of get the sense of what they should be doing in that environment. So, they are no longer in this space where, you know, you just, like, arrive in …”
Hausman: Physical Intelligence's robot demonstrations are fully end-to-end
“So end-to-end robotics is already here. Everything we've shown so far is fully end-to-end where you take camera input in and view other sensors and output actions directly.”
Hausman: End-to-end learning is the only viable path to solve robotics
“There's another reason to do end-to-end learning, which is, this is, I think, the only thing that has a chance of working.”
Hausman: Physical Intelligence demonstrated previously impossible tasks like laundry folding
“The demonstrations that we, that we've shown so far here at physical intelligence are of tasks that were not possible before, like things like folding laundry. You can't really, there is no program that I've ever seen that could do that.”
Groom: Subscale industry gives US time to build domestic robotics supply chains
“Most of the money is being spent on R and D rather than scale production. And so it's not as if there's a 100,000 robots that everyone's buying and it's now just twice as expensive. I think the good thing is that given it's subscale, there's a lot of time to b…”
Hausman: Robotics AI does not yet possess LLM-style compute scaling laws
“We are not there yet in terms of like having a full scaling law the same way as we've seen for LLM companies where you can just translate prog compute to progress to capability.”
Hausman: A small robot fleet generates LLM-scale model training data volumes
“I think that's one thing that, that I realized since starting the company is that robots generate a ton of data and you don't need that many to generate data that is close to the levels that LLM companies use for their models.”
Hausman: Most future frontier AI models will train on real-world robot data
“So I think over time, it's quite likely that, that the places are gonna switch a little bit, where most of the models, including, you know, LLMs and BLMs, are gonna be using real-world data collected through robots, because that's the data that has no ceiling,…”
Groom: Physical Intelligence prioritizes manipulation tasks over conversational UI
“It's not a big focus of ours right now, really. We're so focused on, on manipulation and economically valuable tasks and more so than that, the fundamental building blocks that we think gets us from here to physical intelligence.”
Hausman: Robot physical actions function as another language for multimodal models
“And what we start to realize is that all of these different data sources contribute to each other. They give you just like a bigger picture of what the world is like and better understanding. And it just turns out that robot actions is just like yet another la…”
Hausman: Physical Intelligence's model converses as well as open-source VLMs
“The model that we have already is the model that you can talk to, and it works, you know, just as well as open source BLMs.”
Groom: Waymo is far better than Tesla despite having fewer cars
“Tesla has this, Incredible advantage with how much data they're collecting and passively, yet Waymo is so much better so far, and it has so many fewer cars on the roads.”