Rise of Robots

🧠 Insights You Won’t Forget

Today's insights are inspired by the robots in the Colossus Review

1. Software, Not Hardware, is the Bottleneck

   The key limitation to robot commercialization isn’t mechanical, it’s software. Robotics hardware has advanced and become cheaper, but robots still struggle to function reliably in open, unpredictable environments due to limited reasoning and adaptability software.

2. The Real Frontier is Open Worlds

   The progression in robotics is from “constrained” (factories) to “structured” (warehouses, hospitals) to “open” worlds (homes, sidewalks). Open-world functionality is the ultimate prize, but also the hardest due to its unpredictability and the need for generalist reasoning.

3. Foundation Models for Robots Are Coming, But Need Data

   Just like LLMs needed Common Crawl, robotics needs massive, real-world datasets. Current datasets are tiny in comparison. The Open X-Embodiment project is 1,000 times smaller than datasets used for LLMs. Data collection is the gating factor for training generalist robot models.

4. iRobot’s Roomba Masterclass

   Roomba succeeded not with frontier tech but by solving a clear problem affordably. It was 10x cheaper than competitors and sold 1M units in two years by choosing simplicity, a clear value prop, and leveraging unrelated revenue (military contracts) to sustain early scaling.

5. Sidewalk Delivery Robots Failed Despite the Perfect Storm

   Despite a pandemic boom in food delivery and empty sidewalks, robots captured <0.1% of the market. Why? Sidewalks are “open worlds” filled with unpredictable corner cases. Lack of robust software and real-world data crippled commercialization.

6. Humanoids Are the Formula 1 of Robotics

   Humanoids (like Digit and Figure AI’s robots) are unlikely to commercialize fast, but their R&D pushes boundaries across the robotics stack, balance, motion control, power systems. These breakthroughs may trickle down to more practical robots.

7. Actionable Investment Framework: Three Winning Software Strategies

   • Crack the Data Problem (e.g., Physical Intelligence, Skild AI)

   • Push Structured World Use Cases (e.g., warehouses, hospitals)

   • Forego Autonomy via Teleoperation (e.g., surgical or defense robots)

   Each provides a different way to overcome the limitations of current AI and monetize earlier.

8. Pilot Testing ≠ Product-Market Fit

   In robotics, pilot programs don’t guarantee scalability. Real-world deployment is needed to collect data and refine behavior. This differs from SaaS, where pilot success often predicts scalable revenue.

9. Simulation Helps, But Can’t Replace the Real World

   Simulation (e.g., NVIDIA Isaac) accelerates training, but doesn’t solve the “Sim2Real” gap. Robots still fail on real-world variables like slippery floors or cluttered kitchens. Real deployments are necessary to close this gap.

10. Lessons from Autonomous Vehicles Apply

    The AV industry shows what’s ahead for robotics: decades-long, capital-intensive, data-driven development. Waymo needed 20M miles and Tesla needed billions of customer miles to reach reliability. Generalist robots likely face a similar curve.

Recall from last week

1. China is building the next paradigm: programmable electricity

   Colin makes a compelling case that China’s decade-long investment in electric infrastructure (e.g., AI-managed grids, EVs as mobile power stations, high-voltage lines) positions it as the lead actor in the next general-purpose revolution, akin to the US in 1870.

2. Manufacturing, not software, determines geopolitical power

   The myth of software dominance has run its course. Colin shows how China’s mastery of “process knowledge” and integrated supply chains has positioned it ahead of the US, which is now structurally constrained by short-termism and financialization.

💡 Eko Worth Remembering

⚡ Active Recall – Test Yourself 

🛤️ Off the Record

Finally got around to reading about the robots! Saving my full thoughts for later.

Small announcement: Going to only be writing this ‘Off the Record' section MWF as I have been tighter on time and do not want to sacrifice quality. I appreciate all of your reading 🙂 

Eko’s Top Pods

Answer:

It shows that even seemingly simple environments can be “open worlds” with high unpredictability, requiring sophisticated real-world data and robust software to succeed. Capital and demand aren’t enough without adaptability.

Enjoyed these insights? Forward this newsletter to a friend. Let’s grow smarter, together.